İSTANBUL TECHNICAL UNIVERSITY INSTITUTE OF SCIENCE AND TECHNOLOGY Ph.D. Thesis by Sanem SARIEL, M.Sc. Department : Compu ter Engin eeri ngProgramme: Computer EngineeringJUNE 2007AN INTEGRATED PLANNING, SCHEDULING AND EXECUTION FRAMEWORK FORMULTI-ROBOT COOPERATION AND COORDINATION
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
I first came to focus on distributed intelligence and multi-entity cooperationin 2002, and I am always impressed by the way a group or communityconstitute a great power morally or physically and get around difficulties byacting together, as we observe in nature and in our daily affairs. Working onrobots and investigating of intelligence, particularly distributed intelligence, havebeen my research objectives since my early years at Istanbul Technical University.
Prof. Tucker Balch helped make many of my research dreams come true. Hehas my deepest appreciation for believing in my research, continually supporting
me, and advising me in the way I could improve myself. It has been a pleasureto work with him in the BORG Laboratory at Georgia Institute of Technology.
In conducting this PhD research and writing this thesis book, I have been veryfortunate to have Prof. Nadia Erdogan as my co-advisor. I’m deeply indebtedto her for helping me advance my research. In many cases her affectionateencouragement helped me keep my motivation up.
I feel lucky to have met many great scientists, especially those who greatlysupport and encourage young researchers in their research. I am grateful to Prof.
Martin Savelsbergh, Prof. Pınar Keskinocak, Prof. Sven Koenig, Prof. MichailLagoudakis and Prof. Ashok Goel for their valuable comments on my research.
I would also thank Prof. Emre Harmancı, Prof. Levent Akın, Prof. TevfikAkgun, Prof. Bilge Gunsel and Prof. Turgay Altılar for the continuous moralsupport they provided during my research studies at Georgia Tech.
Many thanks to young geniuses: Ram Ravichandran, Keith O’hara, Arya Irani,Victor Bigio, Ananth Ranganathan, Adam Feldman, Michael Kaess, MattPowers, Sang Min Oh and all the members of the robotic-discussion-group for
the valuable discussions on robot research and for their friendship.
Pınar Taskıran, Sibel Yaman, Faik Baskaya, Can Envarlı and Ping Wang havealways been very supportive especially when I felt lonely during this long journey.My special thanks go to them.
I specifically thank Jacquelyn Berry and Zuhal Yılmazer for their precious effortsin official tasks and for their valuable friendship.
I am deeply grateful to A. Cagatay Talay for his friendship, continuous supportand encouragement throughout this research and his willingness to take over allmy administrative burdens while I was in Atlanta.
I would finally thank my mother, Necla Sarıel, and my sister, Selen Sarıel. ThisPhD thesis research could not have been conducted and completed without theircontinuous support, love and patience.
Finally, this PhD book has come to a conclusion. However, I would like to
emphasize what my advisors always mention:“This is not the end of my research. It’s just the beginning...”
1. INTRODUCTION 11.1. Open Issues in Multi-Robot Coordination Research 21.2. A Brief Overview of the Proposed Approach 21.3. Contributions of the Thesis 3
1.3.1. DEMiR-CF 31.3.2. Formulation of the Cooperative Mission Achievement and
Coordinated Task Selection Problems 5
1.3.3. Integration of Task Allocation, Execution and ContingencyHandling into a Single Framework 5
1.3.4. Real-world Suitability with Limited Assumptions 61.3.5. Investigation and Generation of Novel Solutions for Different
Application Problems 61.3.6. Combination of The Methods from Different Disciplines 6
2. PROBLEM STATEMENT AND MOTIVATION 72.1. Multi-Robot Mission Achievement 72.2. Real-Time Issues and Requirements for Multi-Robot Task
Achievement 9
2.3. Multi-Robot Cooperation vs. Coordination 102.4. Formulation of the Cooperative Mission Achievement Problem 10
3. BACKGROUND AND RELATED WORK 133.1. A Brief Review on the Classification of Earlier Multi-Robot Systems 133.2. Organization and Control Hierarchy 14
5.1.2.3. Heuristics 585.1.3. Robotic Research Methods for the MTRP 58
5.1.3.1. Prim Allocation Method 595.1.3.2. Other solutions for the MTRP 60
5.1.4. Cooperation Objectives 60
5.1.5. Application Domains for the MTRP 615.1.6. Communication Requirements 635.1.7. Formation of the Mission Structure for the MTRP 63
5.2. Applying DEMiR-CF to the MTRP 645.2.1. The Dynamic Priority Based Task Selection Scheme 64
5.2.1.1. Selecting is Eliminating the Other: Incremental Allocationthrough Unconditional Commitments 64
5.2.1.2. Cost Estimation and Evaluation 655.2.1.3. Dynamic Task Selection and Distributed Target Allocation 66
5.2.2. Failure Detection and Recovery 68
5.3. Bounds on the Solution Quality 695.3.1. Analysis of the Approach 695.4. Implementation Details 705.5. MTRP Experiments 71
5.5.1. Experiment 1: Evaluation of the FAC Heuristic Cost Function 735.5.2. Experiment 2: Performance Comparison of DEMiR-CF and the
Prim Allocation Approach with the Optimum 735.5.3. Experiment 3: Real-Time Dynamic Simulation Experiments 765.5.4. Experiment 4: Real-Time, Real-World and Dynamic
Environment Experiments 805.6. Summary and Discussion 84
6. EMPIRICAL EVALUATION OF DEMiR-CF ON NAVYMISSIONS 866.1. Naval Mine Countermeasure Missions 866.2. Applying DEMiR-CF to the Naval MCM 87
6.2.1. Task Representation for The MCM 876.2.2. Exploration for Detection and Classification of MLO Locations 886.2.3. Identification of Mine-like Objects 89
6.3. Experimental Results on the Naval MCM 906.4. A NAVY Homeland Security Application 94
6.5. Summary and Discussion 98
7. EMPIRICAL EVALUATION OF DEMiR-CF ON RESOURCECONSTRAINED AND INTERRELATED TASKS 997.1. Complex Multi-Robot Mission Problem Statement 997.2. Applying DEMiR-CF to Complex Missions 100
7.2.1. Dynamic Priority-based Task Selection Scheme and OnlineScheduling 100
7.2.1.1. DPTSS Algorithm 1037.2.2. Distributed Task Allocation Scheme 1047.2.3. Cost/Bid Evaluation and Tie Breaking Rules 1047.2.4. Analysis of the Approach 105
Table 4.1 Robot Team and Capabilities in the Box Mailing Mission 34Table 4.2 Representation of the tasks of the Box Mailing Mission 35Table 4.3 Message types in DEMiR-CF 46Table 4.4 Precautions for contingencies and conflicts 52
Table 4.5 Model checking for tasks and system robots 53Table 4.6 Model updates related to the messages 53
Table 5.1 FAC heuristic function performance results for the knownTSP instances 72
Table 6.1 The performance results for different message loss rates 94Table 6.2 The cost evaluations for the application domain 97
Figure 3.1 A sample mission representation as an AND/OR task tree 16Figure 3.2 A sample multi-agent plan representation as task graphs 17Figure 3.3 Contract Net Protocol 24
Figure 4.1 DEMiR-CF Modules 32
Figure 4.2 Directed acyclic mission graph for the Box Mailing Mission 33Figure 4.3 The Box Mailing Mission initial state 33Figure 4.4 Basic steps of an auction negotiation process 44Figure 4.5 An invalid auction cancellation 45Figure 4.6 Coalition maintenance 48Figure 4.7 States of the FSMs for single-robot task models 50Figure 4.8 States of the FSMs for multi-robot task models 51
Figure 5.1 The optimal solution can be obtained by clustering the targets 55Figure 5.2 Clustering the targets does not necessarily result in the
optimal solution 56
Figure 5.3 MSF allocations 57Figure 5.4 Effects of the MST traversal strategy on the total cost for
the open traversal version of the TSP 59Figure 5.5 Two different optimization objectives for the MTRP 61Figure 5.6 Target selection strategy by the FAC Heuristic function 66Figure 5.7 The Archimedean spiral and two simple implementations 70Figure 5.8 The robot with correct localization finds the target 70Figure 5.9 An implementation of the Archimedean spiral with the
corresponding robot behavior 71Figure 5.10 Results for the ATT48 TSP instance 72
Figure 5.11 Runtime comparison of the approaches for single robot routegeneration 73Figure 5.12 Performance results of the heuristic approaches 74Figure 5.13 Allocations generated by each approach for an instance of
15 robots and 50 targets 75Figure 5.14 Scalability analysis for different number of robots 76Figure 5.15 Performance results for the Contingency Handling Mechanism 77Figure 5.16 The initial configuration of Scenario 1 78Figure 5.17 The final configuration of Scenario 1 78Figure 5.18 Maps and paths for Scenario 1 79Figure 5.19 The final configuration of Scenario 2 79Figure 5.20 Maps and paths for Scenario 2 79Figure 5.21 The Khepera II robot visits six targets in the environment 80
Figure 5.22 Three Khepera II robots divide the labor of visiting the sixtargets 80
Figure 5.23 The team handles the failure of the third robot in real-time 81Figure 5.24 Single robot cases for different initial deployment locations
of the robots 81
Figure 5.25 Scenario 1: the multi-robot case without failure 82Figure 5.26 Scenario 2: the multi-robot case in which the third robot
fails after completing its task 83Figure 5.27 Scenario 3: the multi-robot case in which the third robot
fails before completing its task 83Figure 5.28 Scenario 4: the multi-robot case in which the first robot fails
before after completing its task 84Figure 5.29 Scenario 5: the multi-robot case in which the second robot
fails before completing its task 84
Figure 6.1 Conceptual flowchart related to the AUV operations 90
Figure 6.2 NAVY MCM Mission Scenario 1 91Figure 6.3 NAVY MCM Mission Scenario 2 92Figure 6.4 NAVY MCM Mission Scenario 3 93Figure 6.5 Initial mission graph consists of only the search task 95Figure 6.6 Robots patrol the area in the corresponding regions 95Figure 6.7 A sample execution trace under highly dynamic situations 96Figure 6.8 Mission graph and allocations evolving through time
accordingly 97Figure 6.9 Execution conflicts under limited communication are resolved 98
Figure 7.1 Mission completion time results for the pick-up/delivery
mission 106Figure 7.2 Total path length results for the pick-up/delivery mission 107Figure 7.3 Scenario 1 and 2 108Figure 7.4 Khepera II robots achieve the overall complex mission 109Figure 7.5 Real scenario mission graph with interrelated tasks 109
Figure B.1 ATT48 TSPLIB instance with 48 nodes 127Figure B.2 Optimal open-loop route for the ATT48 TSPLIB instance 127Figure B.3 EIL51 TSPLIB instance with 51 nodes 128Figure B.4 Optimal open-loop route for the EIL51 TSPLIB instance 128Figure B.5 BERLIN52 TSPLIB instance with 52 nodes 129
Figure B.6 Optimal open-loop route for the BERLIN52 TSPLIB instance129Figure B.7 EIL101 TSPLIB instance with 101 nodes 130Figure B.8 Optimal open-loop route for the EIL101 TSPLIB instance 130
tsj : The most suitable task to execute for robot r j
T : Set of the tasks of the overall missionT φ : Set of all ineligible tasksT Aj : Set of all active tasksT Cj : Set of all critical tasks for robot r j
T Ej : Set of all eligible tasksT Ij : Set of all inactive tasksT Rj : Rough target set for r j
T TSP : Partial tour for the TSPtype : Type of task ti
wi : Waypoint indexed by i
wb : Boundary waypointW jk : Rough Region Schedule for robot r j and region k
FRAMEWORK FOR MULTI-ROBOT COOPERATION andCOORDINATION
SUMMARY
Although several architectures have been proposed for multi-robot coordinationpreviously, the field is still in its early stages and robot teams are still testsubjects in research laboratories. Fortunately, a growing community has beenresearching techniques for multi-robot systems. The basis has been formed
for multi-robot systems to serve humanistic needs in the new future in manydomains, such as search and rescue operations and space explorations in the realworld. In fact, even though some of the open problems have been solved, someopen issues still remain even for single bodies of robots.
The real dynamics of physical task performance force unplanned actions to betaken. Since the world is beyond the control of robots and changes continuouslyin real-world applications, the difficulty of the multi-robot task executionproblem goes beyond the task allocation problem. In particular, multi-robotsystems deal with difficulties arising from noisy sensor information, unexpectedoutcomes of actions, environmental limitations (especially in communication)
and the presence of failures of hardware. Furthermore, the problem instance maybe evolving through real time which is another important research challenge. Allthese factors may affect the overall solution. Against this background, researchin this thesis addresses issues of real-time execution when managing an overallteam by a central authority is not possible due to limitations of the real-worldenvironments. Therefore, each individual robot should find a way to solve theglobal problem from a local perspective in a decentralized way.
The main contribution of this PhD research is the design of a general framework,Distributed and Efficient Multi-Robot-Cooperation Framework (DEMiR-CF),
that can be used to solve problems on a wide variety of application domains.DEMiR-CF is suitable for multi-robot teams cooperating to achieve a globalmission. Team members cooperate to fulfill a mission by dividing the laborof task execution through individual decisions that coordinate their actions,contributing to the achievement of the goal in a distributed manner.
The cooperative mission execution problem is formulated as the CooperativeMission Achievement Problem (CMAP ) and each individual robot contributesto solve the CMAP by solving the formulated Coordinated Task SelectionProblem (CTSP ) for itself. These two problem formulations are introduced to
represent the generalized problem and stated formally for the first time in thisPhD thesis. DEMiR-CF is capable of resolving the CMAP and CTSP for robots.
In the design of DEMiR-CF, the following issues were particularly investigatedas the design criteria: efficient and realistic representation of missions, efficientallocation of tasks to cooperatively achieve the global goal, maintenance of the system coherence and consistency by the team members, detection of the
contingencies and recovery from various failures that may arise during runtime,efficient reallocation of tasks (if necessary) and reorganization of team members(if necessary).
DEMiR-CF is designed to address different types of missions from the simplest tomore complex ones, including missions with interrelated tasks and multi-resource(robot) requirements. In the proposed mission representation, componentsof individual tasks are also allowed to be updated by robots depending onreal-world requirements.
The framework ensures an efficient way of integrating task planning, allocationand execution for multi-entity (agent/robot) teams independent of theunderlying low-level behavior architecture, yet without ignoring it. Theintegrated components of the framework ensure solving the CMAP in a robustand efficient way. As a result, global planning, scheduling and execution iscarried on by the cooperative work of robots performing under DEMiR-CF.Even though an incremental task selection approach is adopted, a global planconsideration is also preserved to make the system both sound and complete.
The performance evaluations of the framework were implemented both onsimulations and on real robots for different application domains. Eachapplication domain is a separate problem domain which requires in depthresearch.
The Multiple Traveling Robot Problem (MTRP) to explore several targets isthe first application domain on which tests were performed. This domain formsa basis for different application domains. Although tasks of this domain areindependent of each other, there is a combinatorial structure of the problemwhen efficiency of the solution is concerned. Therefore, optimization of thegenerated solutions is investigated. Some heuristic cost functions to solve theCTSP is proposed to be used with DEMiR-CF in the thesis. The performance
of the proposed heuristics and the framework is compared with that of one of the well-known allocation approaches.
The evaluations on NAVY domains were performed where cooperation of underwater vehicles is achieved for homeland security missions, such as minecountermeasure missions. In this domain, the robot team is modeled as aheterogeneous team and the mission is constructed from different types of tasks,where each one needs to be performed by a different vehicle. The domain ismodeled as containing both the coverage problem and the MTRP in itself.Robustness of the framework against both communication and robot failures
and efficiency of its response to the dynamically varying conditions are tested insimulations.
In general, DEMiR-CF targets complex missions involving tasks with resourceconstraints and interrelations. Therefore, these types of missions are perfectcandidate domains to apply the full functionality of the framework. In thelast experimental setup, specifically, pick-up/delivery and object construction
domains are investigated as complex domains. Different from previousexperiments, the robots are involved in more complex tasks where they interactwith the objects in the environment. The objective is not only optimizing costfunctions but also obeying rules and resolving constraints on task executionduring runtime. The base mechanisms of DEMiR-CF are used to design thesolution. The efficiency of the proposed solution for complex domains is evaluatedthrough experiments.
Literaturde coklu-robot koordinasyonu icin bircok mimari onerilmis olmasınaragmen, bu konudaki arastırma alanı henuz gelismekte olup uzerinde calısılmasıgereken bircok acık nokta bulunmaktadır. Coklu-robot sistemleri henuzlaboratuvarlarda arastırma denekleri olarak kullanılmakta olsalar da yakıngelecekte bircok uygulama alanı icin insanlık yararına hizmet etmek uzereilk adımlar atılmıstır. Bu uygulama alanlarının basında arama-kurtarmaoperasyonları ve uzay arastırmaları gelmektedir. Gittikce artan bir ilgiylegenisleyen coklu-robot sistemleri arastırma grupları, coklu-robot sistemleri icincesitli metodlar uzerinde arastırma yapmaktadır. Ancak, tek robotlu sistemlericin bile henuz cozulmemis bircok problem bulunmaktadır.
Gercek dunyada ortaklasa otonom olarak calısan donanımların gorev basarımıdaha onceden planlanmamıs davranıslar ve olayları goz onune almayı gerektirir.Robotlar kendilerinin kontrol altında tutamadıkları ve surekli dinamik olanortamlar ve uygulama alanlarında calıstıklarından, coklu-robot gorev yurutme
problemi gorev paylasımı probleminin de otesinde cok daha zor boyutlaratasınır. Ozel olarak, coklu-robot sistemleri gurultulu sensor verileri, beklenmeyendavranıslar ve sonucları, ortamsal kısıtlamalar (ozellikle iletisimde) ve sıkolabilecek donanım bozulma ve hataları gibi cesitli durumlardan kaynaklananzorluklar ile yuzlesmek zorundadır. Bunlara ek olarak problemin kendisi de ortamdurumu veya baska bir sebeple degisim gosteriyor olabilir ki bu da problemidaha zor kılmaktadır. Tum bu faktorler sistemin genel basarımını etkiler. Budoktora tezi arastırması, tum bu durumlara karsı gercek zamanlı gorev yurutmedurumlarını da goz onune alarak ortamdan kaynaklanan kısıtlamalar sonucurobot sistemini tek bir merkezden yonetmenin mumkun olmadıgı durumlar icin
gecerli yontemler uzerine odaklanmıstır. Dolayısıyla her bir robot tum problemiyerel olarak dagıtılmıs bir sekilde cozmek icin bir yol bulmalıdır.
Bu tezin en onemli katkısı, bircok uygulama alanında kullanılabilecek olan genelbir mimarinin -DEMiR-CF (Distributed and Efficient Multi-Robot-CooperationFramework)- tasarımı ve ele alınan uygulama alanları icin bu mimari kullanılarakcozumler onerilmesidir. DEMiR-CF karmasık bir ana isin basarılması icinrobotların ortaklasa calısmasını ongoren bir mimaridir. Ortak calısma,robotların dagıtılmıs olarak karar alması ve gorev paylasımı yaparak butun isinbasarımına katkıda bulunması ile saglanır.
Ortak is yurutme problemi -CMAP (Cooperative Mission AchievementProblem)- her bir robotun bu butun problemi cozmek uzere kendisi icin koordineligorev secim problemini -CTSP (Coordinated Task Selection Problem)- cozmesiile gerceklenir. Bu problem tanımları ve formulasyonları genellestirilmis ortakgorev yurutme problemini temsil etmek uzere ilk kez bu tezde ortaya konmustur.
Onerilen mimari de bu problemleri cozmek uzere tasarlanmıstır.
DEMiR-CF tasarımında ozellikle su problemler ve cozumleri uzerine calısılmıstır:gorevlerin etkin ve gercekci sekilde temsili, gorevlerin genel amacı gereceklemekuzere etkin sekilde atanması, sistem butunlugunun ve tutarlılıgının robotlartarafından korunması, robotların yurutme zamanı olusabilecek olagan dısıdurumlar, hatalar ve bozulmaları tespit edip uygun hata kotarma yontemlerinigerceklemesi, gerekiyorsa gorevlerin etkin sekilde yeniden atanması ve gereklidurumlarda takım uyelerinin yeniden organize olmaları.
DEMiR-CF en basit islerden en karmasık islere kadar farklı tipte gorevleruzerinde calıstırılabilecek sekilde tasarlanmıstır. Karmasık isler birbirlerinebagımlı ve coklu kaynak (robot) gereksinimi olan gorevler icerebilir. Onerilenmimaride, gorevler, yurutme sırasında gercek dunya gereksinimlerini karsılayacaksekilde guncellenebilmesine acık sekilde temsil edilmektedir.
Mimari, coklu-robot (etmen) sisteminin gorev planlaması, ataması ve yurutmesinialt katman davranıs modelinden bagımsız olarak fakat onu gozardı etmeyeceksekilde gercekleme imkanı sunar. Mimarinin tumlesik bilesenleri CMAP
problemini hataya karsı dayanıklı ve etkin sekilde cozmek uzere tasarlanmıstır.Dolayısıyla genel planlama, gorev ataması ve yurutmenin es zamanlı sekildegerceklenmesi, DEMiR-CF mimarisinde tasarlanmıs robotlar tarafından mumkunolmaktadır. Robotlar dinamik ve artımlı olarak gorev secimini yuruturler, ancaktum isi etkin olarak tamamlayacak sekilde genel plan degerlendirilmesi yapılır.
Mimarinin basarımı hem benzetim ortamları, hem de gercek robotlar uzerinde,farklı uygulama alanlarında sınanmıstır. Aslında, her bir uygulama alanıderinlemesine calısma gerektiren ayrı bir problemdir.
Coklu-robot coklu-hedef ziyaret etme problemi, basarım analizinin yapıldıgıilk uygulama alanıdır. Bu problem bircok uygulama alanı icin bir temel
olusturmaktadır. Problem gorevleri birbirlerinden bagımsız olsa da, cozumkalitesi goz onune alındıgında, problemin kombinasyonal bir yapısı vardır.Dolayısıyla uretilen cozumlerin eniyilemesi uzerinde durulmaktadır. Bu tezdebu problemi cozmek uzere DEMiR-CF mimarisinde kullanılabilecek maliyethesap fonksiyonları onerilmektedir. Onerilen bu hesap fonksiyonları mimariile birlikte gerceklenmis ve literaturde iyi bilinen bir gorev atama yontemi ilekarsılastırılarak basarım analizi yapılmıstır.
Bir baska uygulama alanı olarak mimari, kıta sahanlıgını korumaya yonelikaskeri Deniz Kuvvetleri problemleri uzerinde gerceklenmistir. Buradaki robotlar
sualtı aracları olup, ornek bir uygulama problemi olarak sualtı mayın tarama veimha problemi ele alınmıstır. Robotlar farklı yeteneklere sahip olup heterojen
birimler olarak modellenmistir. Ana is, farklı yetenekler gerektiren ve bunedenle farklı robotların yurutmesi gereken alt gorevlerden olusmaktadır. Buproblem icinde, hem ortam tarama, hem de “coklu-robot coklu-hedef ziyaretetme” problemi irdelenmistir. Ayrıca, mimarinin iletisim problemleri ve robotbozulmalarına karsı hataya dayanıklılıgı ve ortam dinamizmine karsı basarım
iyilestirme yetenekleri irdelenmistir.
DEMiR-CF mimarisi, en genel anlamda coklu kaynak gereksinimleri olan vebirbirleri ile baglantılı gorevlerden olusan karmasık isleri cozmeyi hedef alır.Dolayısıyla bu tur karmasık gorevler, mimarinin tum fonksiyonelligini olcmekuzere en uygun aday uygulama alanlarını olusturur. Son deneysel platformda,karmasık gorevler olarak nesne alma/tasıma ve nesneler ile insa uygulama alanlarıuzerinde calısılmıstır. Onceki deneylerden farklı olarak, robotlar ortamdakinesnelerle de etkilesim kurdukları karmasık gorevler uzerinde calısmıslardır.Amac sadece eniyileme degil, aynı zamanda kısıtlama ve kurallara uygun
sekilde gorev yurutmektir. Bu uygulama alanlarında, DEMiR-CF mimarisinintemel bilesenleri kullanılmıstır. Gerceklenen deneylerde yontemin etkinligiirdelenmistir.
Sometimes robot application domains contain assemblage of these missions.
Some of these domains are adversarial, that is, robots compete against each
other, such as in multi-robot soccer teams. However, in this research, we focus
on cooperative multi-robot systems. Our proposed approach may also fit in
adversarial domains as well, interpreting the competition as facing difficulties of the environment in a cooperative domain.
1.1. Open Issues in Multi-Robot Coordination Research
Although several architectures have been proposed for multi-robot coordination,
the field is still in its early stages with a great deal of open questions in many
dimensions. In fact, even though some of the open problems have been solved
in single robot research, some open issues still remain even for single bodies of
robots.
Open issues of multi-robot coordination that we would like to address are:
• How can tasks be represented for effective cooperation?
• Which robot should perform which task in order to cooperatively achieve
the global goal? Who should take which role in the team?
• How is group coherence and consistency maintained?
• Which information is shared when there are environmental, communication
and/or process power limitations?
• How are contingencies detected?
• How are allocations re-organized when contingencies are detected?
In addition to these open questions, Parker (2004) brings forward the open issue
of “whether a general architecture can be implemented to cover a wide area of applications” in her multi-robot research review.
1.2. A Brief Overview of the Proposed Approach
This PhD research aims to focus on the investigation of the open issues in
multi-robot task allocation and execution problem in the context of autonomous
achievement of a mission and to propose novel algorithms to address open issues,
explicitly for task allocation and execution. We do not touch on research onlow-level essential robotic elements such as localization, mapping, etc. Our
Our research addresses issues of real-time execution when the managing of the
overall team by a central authority is not possible due to limitations of the
real-world environments. Therefore, each individual robot should find a way
to solve the global problem from a local perspective while thinking as globally as
possible as in the third allocation scheme.
2.2. Real-Time Issues and Requirements for Multi-Robot Task
Achievement
Even in the case of carefully written orchestra scores or playbooks, the real
dynamics of physical task performance force some unplanned actions to be taken.
Since the world is beyond the control of the robots and changes continuously in
real-world applications, the difficulty of the multi-robot task execution problem
goes beyond the task allocation problem. In particular, multi-robot systems deal
with difficulties arising from noisy sensor information, unexpected outcomes of
actions, environmental limitations (especially in communication) and the presence
of failures of hardware. All these factors may affect the overall solution. We list
evolving circumstances that may change the solution as:
• Self failure detection: Robots detecting their own failure.
• Robots detecting the failure of another robot.
• Change in the estimated task execution cost/time: Environmental
dynamics, uncertain knowledge, or hardware problems may cause delays
in task execution or early achievements of tasks. Uncertain sensor and/or
localization information may also result in incorrect estimations.
• Change in the task definitions: Task dependencies, priorities, or the overall
objective (goal) may change. Some tasks may become invalid during
runtime.
• New online tasks introduced by human operators or discovered by the robots
themselves.
• New robots being released, or some failed robots being repaired or
recovering from trap-like threats.
• Intervention and manual changes on assignments by external agents.
Some of these situations may arise after either internal or external events. Given
these contingencies, even the result of an approach capable of finding the optimalsolutions may become suboptimal under the uncertainties of the real-world
The goal of this section is to introduce the main concepts and basic vocabulary
needed to comprehend the proposed approach. The material has been classified
under different subtitles including corresponding related work to give better
insight into the principles used in the research. Each corresponding section first
introduces coordination mechanisms, then reviews the earlier implementations
that utilize them.
3.1. A Brief Review on the Classification of Earlier Multi-Robot
Systems
There are two different types of approaches for multi-robot coordination:
implicit/emergent coordination and explicit/intentional coordination. Implicit
coordination methods take their grounds from biological inspirations. These
systems are suitable for large teams to effectively achieve an overall mission
with the local view of each individual team member. This approach produces
highly effective solutions when combined with behavior based architectures fordomains such as foraging (Balch and Arkin, 1998). However, these systems are
usually domain-dependent. Explicit coordination schemes, on the other hand,
embody more complex representations and algorithms compared to the implicit
case, so their generalization is much easier. They may be extended to address a
wide variety of applications if implemented effectively. In this section, we review
the existing work in explicit multi-robot coordination for team tasks in detail,
leaving implicit approaches out of our scope.
Gerkey and Mataric (2004) present a taxonomy for the Multi-Robot Task
Allocation (MRTA) problem. In their analysis of this problem, they state that
utility is the core subject of optimization of overall solution quality. Since utility is
a measure of both the robot’s state and that of the environment, its inexactness
due to sensor noise, uncertainties, and changes in the environment makes the
multi-robot coordination problem difficult. Additionally, it is emphasized that
it is often difficult to measure the main objective being optimized during
execution. In their taxonomy, the problems can be classified within three different
Recent studies have revealed that the distributed approach, especially when
complemented with the auction-based methods, has shown great promise for
multi-robot task allocation during the last decade because of its scalability.
M+ (Botelho and Alami, 1999) is one of the most successful architectures for
distributed task allocation and achievement, addressing many real-time issues,including plan merging paradigms. MURDOCH (Gerkey and Mataric, 2002) is
a framework that achieves publisher/subscriber type allocation for instantaneous
assignment. Dias (2004) proposes a combinatorial auction-based task allocation
scheme: TraderBots. Zlot and Stentz’s work (Zlot and Stentz, 2006) on task tree
auctions is presented as an extension to the Traderbots approach to address more
complex tasks which can be decomposed into task trees. Lemaire et al. (2004)
propose a task allocation scheme for multi-UAV (Unmanned Aerial Vehicles)
cooperation with balanced workloads of robots. Recently Gancet et al. (2005)have proposed a framework for multi-UAV coordination. The application problem
requires consideration of temporal constraints, uncertainties of task execution,
reactivity to contingencies and changing priorities during runtime by operators.
Their approach supports both centralized and distributed coordination and uses
synchronization signals to coordinate synchronization among UAVs. We review
these architectures in the following corresponding subsections.
3.3. Task Representation for Coordination
Tasks need to be efficiently represented before they are allocated among the
robots. This may require that the overall mission be decomposed into several
tasks.
In earlier systems, usually ad-hoc representations are used to represent the
mission structure. Some models for task representations for robots such as
Task Description Language (TDL) (Simmons and Apfelbaum, 1998) exist in
the literature; however, to our knowledge, a TDL representation supportingmulti-robot coordination has not been released or published yet. Goldberg et al.
(2002) use TDL in the executive layer of their layered market-based coordination
architecture. TAEMS framework ensures ways to define interrelations among
tasks in a detailed expression syntax (Decker, 1996) for multi-agent coordination.
Zlot and Stentz (2006) use AND/OR trees (Nilsson, 1986) to represent alternative
and entailed solutions for task execution. In Figure 3.1 a sample task tree
representation can be seen for the reconnaissance mission. In this mission
description, the area should be covered by dividing the region into different parts
(AND connections) in different ways (OR connections).
Figure 3.1: A sample mission representation as an AND/OR task tree (Zlotand Stentz, 2006)
The task dependency issue, on the other hand, presents orderings among tasks
and has not been given much attention in earlier multi-robot systems. The
interdependencies make the NP-Hard multi-robot coordination problem even
harder. The interrelations among tasks require effective task representations. As
Malone and Crowston (1994) state; if there is no interdependence, there is nothing
to coordinate. Lemaire et al. (2004) represent temporal interdependencies as
task trees and the task execution is synchronized among robots by temporary
master-slave relationships.
Graph representations may be more suitable for representing tasks with
precedence relations as in activity on node graphs. This representation enables
the usage of graph algorithms on these graphs. In multi-agent research, each
agent’s abstract plan is given as a separate graph and interrelations are defined
on this graph. A sample multi-agent plan (Cox et al., 2005) for the Multiagent
Plan Coordination Problem (MPCP) is given with interrelations between tasks
(represented as boxes), preconditions and post-conditions in Figure 3.2. Thepreconditions are labeled on the arcs, whereas the postconditions are stated near
the task boxes.
3.4. Task Allocation and Coordination Type Based on the Mission
Structure
We classify task coordination for multi-robot systems into four subclasses, namely,
tightly coupled tasks, loosely coupled tasks, interrelated tasks and tasks as parts
Task dependency has been analyzed in some earlier multi-robot cooperation
schemes. The work by Alami et al. (1998) on multi-robot coordination presents a
generic scheme based on a distributed plan-merging process. Although optimality
is not guaranteed, Plan Merging Operation (PMO) provides a coordinated
plan in their approach. A deadlock resolution is implemented in a distributedmanner. The M+ scheme (Botelho and Alami, 1999) combines local planning
and negotiation for task allocation, and cooperative reaction for contingencies.
In their framework, a mission is a set of partially ordered tasks. Each robot
has its own local world knowledge. Tasks are allocated through negotiation
processes. Alami and Botelho (2001) introduce the mechanism concept in the
framework M+CTA as an improvement to the M+ scheme. Mechanism is used
for the resources in multi-robot cooperation. Each robot has an individual plan
and tasks are initially decomposed and then allocated. After this planning step,robots negotiate with each other in order to incrementally adapt their plans in a
multi-robot context.
3.4.3. Combinatorial Effects on Task Assignment
Tasks with combinatorial structures take part in the task classes with soft
interrelations (i.e., not as hard as in tightly coupled tasks) in our task
classification. In this class, although there is no interrelation among tasks in
the mission definition, because of the combinatorial structure of the problem,
the solution quality highly depends on which task is performed first and by
which robot. Multi-robot exploration as a Multiple-TSP problem, the commonly
studied application domain, belongs to this class of tasks. Separately Lagoudakis
et al. (2004), Dias and Stentz (2002) and Lemaire et al. (2004) have studied this
problem.
3.5. Instantaneous Task Assignment/Scheduling
Multi-robot task allocation is better viewed as a scheduling problem when thereare interrelations and dependencies among tasks. When the problem solving time
is limited, Branch and Bound or MILP methods may not be convenient. In this
case, heuristic methods are preferred to find a good solution in reasonable time.
Furthermore, finding a solution with these approaches in a decentralized setting
may need considerably greater efforts in both computation and communication.
Given these limitations, instantaneous task assignment becomes profitable as
it provides a dynamic solution allocating tasks to robots whenever resourcesare available (Parker, 1998; Gerkey and Mataric, 2002). However, in this case,
the global solution quality may be degraded if the decisions are made by just
using the up-to-date knowledge available, ignoring the global solution quality.
Paquet (2006) models the multi-agent task assignment problem as a scheduling
problem for the RoboCupRescue simulation domain. The main objective is
the maximization of the number of rescued civilians in a simulated disasterenvironment. They compare both centralized and distributed scheduling in their
work. They conclude that the distributed scheduling performance is as good as
that of the centralized scheduling and the decentralized scheduling approach is
more robust. In the distributed approach, each agent locally chooses its best
task to accomplish using a scheduling algorithm. Then, the best local task
information is exchanged among agents to find the global best task to perform.
The Earliest Due Date (EDD) algorithm is used to schedule the tasks both
locally and globally. If there is no overload, this algorithm is optimal for thegiven objective of maximization of the number of rescued civilians. The EDD
algorithm is a useful tool to schedule tasks that have no interdependencies and to
make instantaneous assignments. In the EDD algorithm, future considerations
and interrelations are not taken into consideration.
3.6. Reallocation and Dynamic Task Switching
Depending on the application domain and the frequency of the change inthe world correspondingly, task reallocation is needed to adapt to changing
situations. In dynamic games, such as RoboCup soccer domains, this frequency
is high. However, this frequency may be much lower when robots run in fully
structured environments with known maps and few unexpected events (e.g.,
robots working in a fully structured and isolated factory environment.)
Lagoudakis et al. (2004) propose a task reallocation method to be applied
whenever the world knowledge of the robots changes in the multi-targetexploration domain. On the other hand, dynamic scheduling and task selection
(Paquet, 2006) prevent rescheduling all the tasks that have been previously
scheduled. As stated in Paquet (2006), rescheduling each time that the world
knowledge changes could make agents switch between tasks frequently. In their
work, task preemption is not allowed to circumvent this situation.
In ALLIANCE (Parker, 1998), responding to unexpected events and dynamic
task reallocation are provided through the use of motivations: robot impatience
and robot acquiescence. The impatience motivation enables a robot to handle
situations when other robots fail in performing the given tasks. The acquiescence
motivation enables a robot to handle situations in which it fails to properly
perform its task. In Traderbots (Dias et al., 2004), task reallocation is achieved
through continuous auctioning by one of the robots. In MURDOCH (Gerkey
and Mataric, 2002), task reallocation is provided through the observations anddirectives of the leader robot for tight coordination, and a publisher/subscriber
type of instantaneous task allocation approach for loosely coupled tasks. In
M+ (Botelho and Alami, 1999), contingencies in task execution are handled by
re-planning for execution of the goals at hand by each robot. Chaimowicz et al.
(2002) address the task dependence and role exchange issues in their work. Utility
calculation is implemented to perform role exchanging. Task announcement is
used to call for additional help. In the work by Lemaire et al. (2004), allocations
are implemented whenever the world knowledge of robots changes (there may benew online tasks or robots may fail to achieve their plans).
3.7. Bounds on the Solution Quality
The exact bounds of the solution quality is hard to evaluate for robot systems due
to uncertainties. Therefore, in the last decade researchers have proposed effective
approaches and opportunistic methods without giving boundaries on the overall
solution quality. The one exception is Lagoudakis et al. (2004); however, their
work assumes perfect communication and contingencies are not considered inthese boundaries.
3.8. Organizational Requirements on Multi-Robot Task Execution
and Coordination
Tasks can be classified according to single robot/multiple robot requirements
for task execution, as in the taxonomy given in Gerkey and Mataric (2004).
Furthermore, robots in multi-robot task execution may be part of either a
homogeneous or a heterogeneous group. The heterogeneity may be in the
possessing capabilities or in the task execution performance. For example, the
robots may have the same equipment capable of achieving all the tasks of the
mission but may differ in abilities such as speed. Dahl et al. (2004) address
this issue in their task allocation method through vacancy chains. This method
ensures differentiation between robots based on their individual performances not
related to their physical and sensor capabilities. High-value tasks are assigned
Horling and Lesser (2005) classify the multi-agent organizations depending on
the objective. According to their classification, coalition as an organizational
paradigm is used in systems where the agents form coalitions (agent groups) to
perform a task in cooperation, and the coalition dissolves when the corresponding
task no longer needs to be executed by them. From our perspective, coalitionsare suitable to achieve the tasks which require a subteam to be formed during a
time period.
3.8.1. Coalition Formation
A coalition is an alliance between entities, during which they cooperate in joint
action, each in their own self-interest. This alliance may be temporary or a
matter of convenience. A coalition thus differs from a more formal covenant
(Wikipedia). The selection of the members of a coalition at one step has a greateffect on future coalition formation.
Shehory and Kraus (1998) present one of the earlier algorithms for coalition
formation for cooperative multi-agent systems. During coalition value
calculations, agents’ capabilities are taken into consideration. In multi-robot
systems, the cost values are a function of not only capabilities but also the
physical conditions which change during execution, such as robot’s location,
object/subject’s location, etc. When the robots decide to perform a task, both
subjects in the environment and robots’ physical entities (e.g., fuel) change.
Vig and Adams (2005) state the differences of multi-robot and multi-agent
coalition formation issues from a sensor possessive point of view. Locational
sensor capabilities are considered in their work. They propose an approach
based on the coalition evaluation step in Shehory and Kraus’ algorithm (Shehory
and Kraus, 1998). Vig and Adams’ approach assumes that capabilities are
known apriori before coalitions are formed. This approach may be applicable
in the beginning of mission execution. However, another important factor in
multi-robot systems for evaluating coalitions is the changing cost values during
runtime. They assume robot capabilities do not change (which also is not a
realistic assumption); however, this is not the case for the costs. They analyze
the trade-off between distributing the coalition value evaluation among robots
and implementing coalition value evaluation for each robot. This approach
may result in a robust system where all robots are informed about the robot
failures. Coalition imbalance issue is addressed to make the coalitions more
robust by distributing the required resources. This issue is debatable based onthe objective function which is selected.
ASyMTRe (Parker and Tang, 2006), uses reconfigurable schema abstraction for
collaborative task execution by providing sensor sharing among robots. The
fundamental building blocks of ASyMTRe are collections of environmental sensors
(ES), perceptual schemas (PS), motor schemas (MS), and also communicationschemas (CS). In ASyMTRe, connections among the schemas are dynamically
formed at runtime. The information labels provide a method for automating
the interconnections of schemas, enabling robots to share sensory and perceptual
information as needed. In their approach, initially the schema set is reduced to
contain only the individual separate schemas. Then potential solutions are found.
Finally the corresponding schemas are instantiated on robots. This approach is
used to form low-level coalitions to share robot capabilities.
3.9. Communication and Coordination Tools
Being a part of the real environment, robots trying to achieve common goals need
to interact with others to perform the tasks. Interaction may take place in two
forms:
• by a communication language (direct interaction)
• by observing others (indirect interaction)
Communication is one of the most important coordination tools for robots
when they have joint goals or interacting actions. Some tasks may be
achieved without communication (Balch and Arkin, 1994) but mostly intentional
cooperation/coordination requires some level of communication. Research on
observing other robots is in its early stages with current technology.
Communication protocols are specified at several levels:
• the bottom layer specifies the method of interconnection,
• the middle layer specifies the format, or syntax of the information being
transferred,
• the top layer specifies the meaning, or semantics of the information.
There are two main types of interaction to allocate tasks and reaching a consensus
on task execution: Negotiations and Auctions.
3.9.1. Negotiations
Negotiation is a process by which a joint decision is reached by two or more agents,each trying to reach an individual goal or objective (Huhns and Stephens, 1999,
Muller, 1996). At this point, it differs from cooperative multi-robot objectives.
However, even when the robots work cooperatively, this mechanism can be used
to reach an argument for the global goal from local perspectives.
3.9.2. Auctions
Auctions are inevitable parts of the Market-Based approach which is based on
economy theory. An auction consists of an auctioneer and potential bidders.
Auctioneers want to sell items and get the highest possible price, whereas bidders
want to acquire the item with the lowest possible price (Sandholm, 1999).
The objects of the auctions can either be single items or a bundle of items.
Both single item (Lagoudakis et al., 2004) and multiple item allocation for
multi-robot task allocation (Berhault et al., 2003; Dias, 2004) are studied in theliterature. As a special case of the combinatorial auction approach, the task
tree auctions presented in Zlot and Stentz (2006) propose a way to offer task
allocation from any point in the task tree with the payoff of the expensive bid
clearing algorithms. In the combinatorial auction methods, even when effective
methods are used to define bundles of items, communication requirements and
efforts for negotiations and bidding grow exponentially with the task size.
3.9.3. Contract-Net-Protocol
Even though the issue of task allocation is analyzed in a wide variety of
works, there still are not formalisms to decide which one to implement in a
particular domain. Both single item and combinatorial auctions are proposed for
multi-robot systems. However, there is not a perfect answer for the suitability of
these methods for different domains. Since in many domains the environment is
highly dynamic, the solution to task allocation problem in the initial world state
may be obsolete or suboptimal in later steps. This is due to the environmental
changes after actions of the robots, changes in the robots’ capabilities, changes
in the team mission, or changes in the team capabilities or composition.
Auction-based methods and Contract Net Protocol (CNP) (Randall and Smith,
1983; Smith, 1980) seem to be an efficient way to allocate tasks in a distributed
manner and these methods are applied to both software agents and real robots.
CNP is modeled on the contracting mechanism used by businesses to govern the
exchange of goods and services. The contract net provides a solution for finding
the most appropriate agent to work on a given task. In the contract net protocol,
an agent that has a task to be solved is called the manager. Agents that might be
able to solve the task are called potential contractors. The manager decomposes
its larger problem into a set of subproblems and announces each subproblem to
the network, along with the specifications about which other agents are eligible
to accept the subproblem, when the deadline is reached, and how they shouldspecify a bid for the problem. A recipient of the announcement decides whether
it is eligible, and if so it formulates a bid. The manager collects bids, and awards
the subproblem to the contractors with the best bid. A contractor receives
the subproblem details, solves the subproblem, and returns the solution to the
manager. The contract net protocol procedural steps can be seen in Figure 3.3.
There is a mutual selection between the manager and the contractors. Any agent
The Manager
Decompose
the task
Consider the
announcement
spefications, and
own capabilities
Send a bid
Send a bid
DEADLINE
Choose the
best bid
Award
Solve the
problem
Send the solution
For each subtask
Contractor 1Contractor 2
Contractor N Announce the task
Figure 3.3: Contract Net Protocol
can act as a manager by making task announcements, and any agent can act as
a contractor by responding to the task announcements. The task announcement
can be made by broadcasting to the network or by directed contracting by prior
experience. One drawback of the contract net protocol is that a task might be
awarded to a contractor with limited capability if a better qualified contractor is
busy at the award time. There are also some situations when the manager doesnot get any bid from contractors. The contractors might be busy with other
tasks. They might consider other announcements and prefer them by ranking,
or they might not be capable of working on the task in consideration. If no
contractors are found, the manager may request a response from contractors
indicating their situation such as: eligible but busy, ineligible, or uninterested.
The manager can retry the announcement periodically. The manager canalso relax the eligibility requirements, but there may be no contractors having
capability for the announced subproblems. In this case, the manager tries to
decompose the problem differently. Another solution is that the contractors
can announce availability. It is also possible to alternate between task and
availability announcements.
Although Contract Net Protocol presents the formalism on the relationships
between managers and contractors for task allocation, it does not present detailsfor the following questions:
• What is the communicated information in auctions for multi-robot systems?
• When should task announcement be made? Who should announce the task?
• How should bid values be defined to get globally good solutions for different
domains?
• Which subset (or all) of the already allocated tasks should be re-allocated?
• When should reallocations be announced?
These questions are usually left open in most multi-robot systems.
3.10. Failure Detection and Recovery
Robustness is an important issue for multi-entity systems (Kaminka and Tambe,
2000). The dynamic task allocation problem has been investigated in the face
of robot failures (including partial/complete failures) or environmental changes.In most cases which include failure detection, the test domain missions include
independent subtasks that can be executed by a single robot.
ALLIANCE (Parker, 1998), provide a mechanism to handle robot failures through
using the motivational behaviors. Dias et al. (2004) investigate the performance
of their framework, Traderbots, against different kinds of failures such as
communication failures, partial malfunction or death. In their work, execution
conflicts between robots are resolved by continuous auctioning by one of therobots. Since the robots sell different portions of their tasks during execution,
robot failures are handled by complicated queries on earlier communications
made by the failed robot and communicating with other subcontractors for the
tasks of the failed robot. Gerkey and Mataric (2002) evaluated the MURDOCH
against the robot failures for tightly coordinated task execution in which there
is a leader giving appropriate directives for the changing situation of the systemafter failures. In M+ (Botelho and Alami, 1999), contingencies in task execution
(task failures) are handled by re-planning for execution of the goals at hand by a
robot. The watch-out task introduced in the work by Lemaire et al. (2004) has an
interesting property providing cooperative work against communication failures.
3.11. Interleaving Planning and Execution
According to desJardins et al. (1999), there are three ways to accommodate
planning and execution into one framework:
1. Conditional planning: For each contingency, an alternative course of actions
is provided.
2. Plan monitoring and repair: The plan-and-execute cycle is repeated
sequentially whenever the execution does not match the estimated model.
3. Interleaving planning and execution together: This method corresponds to
continual planning.
As desJardins et al. (1999) state, an agent should plan continually:
• when aspects of the world can change beyond the control of the agent,
• when aspects of the world are revealed incrementally,
• when time pressures require execution to begin before a complete plan can
be generated, and
• when goal objectives change.
desJardins et al. (1999) come up with an important corollary that states it is
better to delay plan refinement as long as possible, so that detailed decisions are
made with as much information as possible. They argue that simply combining
distributed and continual planning methods independently may not be sufficient
and more intelligent integration approaches are needed.
RoboCup domain (Kitano, 2000) is a testbed to develop and improve robot
architectures and algorithms by motivating through competitions held annually.
Vail and Veloso (2003) present a dynamic assignment based coordination
approach for RoboCup soccer robots. Their method is based on shared potential
functions related to the positions of the relevant obstacles. The bidding
mechanism is implemented for distributed coordination. Kose et al. (2005)
present a market-driven coordination approach for RoboCup domain. They
present different bid functions to assign roles and further extend the system
by integrating reinforcement-learning to learn the role assignment process.
Paquet (2006) models the multi-agent task assignment problem as a scheduling
problem for the RoboCupRescue simulation domain. The main objective is
the maximization of the number of saved civilians in a simulated disaster
environment.
As a humanitarian domain the real Search and Rescue (SR) domain is one of
the attractive application domains for deploying multiple robots. Jennings et
al. (1997) propose an algorithm for a distributed team of autonomous mobile
search and rescue robots. In their experimental setup with two robots, the
tasks are explicitly defined and communication is assumed to be perfect. The
experimental robots are homogeneous and they both aim to manipulate anobject with no constraints on the task requirements. Recently, Koes et al. (2005)
propose a centralized architecture with a Mixed Integer Linear Program (MILP)
approach to multi-robot coordination for the search and rescue domain. The
evaluations are performed in simulations.
In the cooperative transportation domain, a group of robots locate and
cooperatively transport several objects scattered in the environment. Chaimowicz
et al. (2002) address the task dependence and role exchange issues in this domain.Dahl et al. (2004) present the results of their approach based on vacancy chains
in this domain. The scheme by Alami et al. (1998) is tested on the multiple
transportation of containers in different environments.
Another attractive domain for researchers is surveillance, monitoring and
reconnaissance domains. These domains could be modeled by either the
multi-robot coverage (Hazon and Kaminka, 2005, Rekleitis et al., 2004) or the
multi-target exploration problem formulation. In multi-target exploration, the
robots visit targets (special observation points or object locations). Goldberg
et al. (2003) present evaluations of their market-based architecture for the
MARS exploration application domain. Dias and Stentz (2002) evaluate the
performance of their framework, Traderbots on the multi-robot exploration
problem in which robots are homogeneous and tasks have the same type of
capability requirements. Zlot and Stentz (2006) evaluate the market-basedcomplex task allocation approach work on the area reconnaissance problem.
The system proposed by Lemaire et al. (2004) aims to be used in surveillance
and monitoring, specifically forest fire monitoring applications. There are other
military applications, such as perimeter sweeping (Kalra et al., 2005) in which
robots try to sweep an area while coordinating their movements to form security
barriers.
The Multiple-TSP problem, the commonly studied application domain, usually isipart of one of the domains presented above. This domain deserves investigation
due to its easy applicability into different applications such as SR and space
exploration domains. Lagoudakis et al. (2004), Dias and Stentz (2002) and
Lemaire et al. (2004) have studied this problem.
3.13. Multi-Robots vs. Multi-Agents
How multi-robot research differs from multi-agent research is questioned by both
robot and software agent researchers. The main difference is in the environmentand the body of the entities; the robots live in the real physical world, whereas
the software agents live in electronic environments. Therefore, the assumptions
made in multi-agent systems make the real robot systems far from reality. While
the agents can travel easily over communication channels, the robots need to
avoid real obstacles while simultaneously localizing themselves and mapping
their environment.
Usually the agents in the multi-agent systems are modeled as self-interested.The agents have their own abstract plans and should coordinate their
activities/actions in a compatible and mutually supporting manner. The
suitability of self-interest in multi-robot systems is a part of ongoing debates.
The main difficulty with robot systems is the requirement for simultaneous
computation and physical actions. The capabilities of the agents are usually
assumed to be shared by other agents. However, considerable effort is needed to
achieve this in robot systems. Simultaneity and synchronization between robots
is much harder with limited sensors and communication capabilities.
The processing ability of each robot is classified as PROC-TME, Turing machine
equivalent. Collective composition is heterogeneous (CMP-HET) where both a
heterogeneous and homogeneous team is addressed.
The overall objective of the robot team (r j ∈ R, 0 < j ≤ ||R||) in our frameworkis to achieve a mission (M ) consisting of independent or interrelated tasks T i
(0 < i ≤ ||M ||), by incremental assignment of all T i ∈ M to r j ∈ R while
optimizing the specified objective function. Tasks are preemptive: The activity
of task execution can be split during runtime if another advantageous situation
arises or environmental conditions impel.
Coalitions (Coali) (Horling and Lesser, 2005) are formed to meet simultaneous
execution requirements of tasks (T i) synchronously by a group of robots. Anexample of such a task that needs to be executed by a coalition of robots is
pushing a heavy object requiring more than one robot. Sizes of coalitions vary
according to the minimum number of robots required (reqnoi) to execute the
tasks. A coalition 1 may involve only one robot for a task that can be executed
by a single robot.
The robots can detect and recover from different types of contingencies by
keeping the models of the system tasks and other robots as corresponding Finite
State Machines (FSM) in their world knowledge. Details of these FSMs are
presented in Section 4.8. The Model Update Module is responsible for checking
and updating a robot’s own models. The modules that embody the framework
and the information which flows among them are given in Figure 4.1. Each robot
keeps a model of the other robots and the mission tasks. The Model Update, The
(System) Consistency Checking Module, and The Dynamic Task Selector Module
perform Plan B Precaution Routines by either updating the model maintained by
the robot or activating the warning mechanisms. Model updates are initiated by
either incoming information from the other robots or information perceived by therobot itself. If a system inconsistency exists, The Consistency Checking Module is
responsible to initiate warning mechanisms and inform the corresponding robots.
The Dynamic Task Selector Module selects the most suitable task by considering
the model of the robot. The Allocation Scheme ensures the distributed task
allocation by executing the required negotiation procedures for the selected task.
The Execution/Coalition Scheme implements synchronized task execution and
coalition maintenance procedures. According to the selected task and the task
1The term coalition is also used for a single robot executing a task for the sake of generality.
corresponding auction and coalition formation procedures (2-4) are applied
continually.
Real-time situations in which task switching becomes necessary are given in
Section 2.2.
Move 1
Move 2
Drop 1
Drop 2
S
T
{Conjunctive Arc}
Clean
[reqno = 2] [reqno = 2]
[reqno = 1]
[reqno = 1]
[reqno = 1]
Figure 4.2: Directed acyclic mission graph for the Box Mailing Mission
Figure 4.3: Box Mailing Mission initial state is illustrated. The robots arelocated on the left. The two boxes, the stamping machine and the mailbox arelocated in different places in the environment.
4.2. Mission Representation
A mission in DEMiR-CF is represented by a directed acyclic graph (DAG) where
each node represents a task and the directed arcs (conjunctive arcs) represent
Example 1 To clarify, Figure 4.2 depicts the graph representation of a small
sized mission il lustrated in Figure 4.3. The mission involves moving boxes to a
stamping machine, dropping them in a given order, and then cleaning the room.
The room can only be cleaned after both boxes are moved. Since box 1 is heavy, two
robots are needed to move and drop the box (hence, reqno = 2). Unique dummy nodes are added to the graph to represent initial (S) and termination tasks (T).
Therefore, even when the task graph is not connected, after adding these task
nodes, it becomes connected. Although this graph shows the relationships between
dependencies among tasks, it does not show which robot performs which task and
in what sequence. The decision of which robot will be involved in the task execution
of a particular task has an important effect on the performance. There may be
alternative solutions to this problem according to the number of available robots,
their capabilities and task requirements. A sample set of robot capabilities is given in Table 4.1.
Table 4.1: Robot Team and Capabilities in the Box Mailing Mission Domain
Table 4.2: Representation of the tasks of the Box Mailing Mission
id type reqcap deplist repno relinfo precinf o
0 M O V E 1, 2 − 2 < locations of object 1 and the final destination > available
1 M O V E 1, 2 S 0 1 < location of object 2 and the final destination > available
2 DROP 1, 2 H 0 2 < location of object 1 > available
3 DROP 1, 2 H 1, H 2 1 < location of object 2 > available
4 CLEAN 3 H 0, H 1 1 < cleaned portion of the environment > available
5. reqno: The minimum number of robots required to execute the task,
determined either before mission execution or during runtime.
6. relinfo: Descriptive information regarding task type, such as the latest
location, the target location, etc.
7. precinfo: Precaution information used for contingency handling: the task
state, the estimated task achievement time and the current execution cost.
Information in a task representation can dynamically be modified during
execution. In particular, relinfo, precinfo and reqno are subject to change
during execution.
Task representation for the tasks of the sample mission described is given in Table4.2. Capabilities required for task execution, such as possessing a bumper (to push
an object), possessing a gripper (to hold and drop a box) and possessing a brush
(to clean the environment) are encoded as 1, 2 and 3, respectively. Soft (S ) and
Hard (H ) dependencies are identified with the corresponding task ids. precinfo
values of the tasks are initialized to state available before mission execution.
These values are updated during runtime. These issues are explained in Section
4.8.
4.3. Inputs and Outputs of DEMiR-CF
DEMiR-CF uses the planning graph of a mission for interrelated tasks represented
as explained. Although planning is assumed to be performed outside the
framework, it is achieved by DEMiR-CF for independent tasks. For example, the
multi-target exploration mission has independent tasks, but the selection of the
targets to be visited by robots is a form of planning which is done by DEMiR-CF.
Another input to the system is the robot capabilities. The system may involve aheterogeneous team of robots both possessing different equipment (e.g., sensors,
different hardware tools, etc.) and/or having different capabilities with the same
equipment (e.g., speed). The robots are informed about the other robots in
the system initially. But this does not preclude that they discover other robots
working on the same mission during runtime in peer-to-peer communications.
The initial information is better to be fed into the robots’ world knowledge,although DEMiR-CF can also handle distributed information update in runtime.
At the end of the mission execution, all achievable tasks of the mission are
performed by robots working in a cooperative manner. In the meantime,
DEMiR-CF also ensures efficiency with its integrated task allocation and
contingency handling capabilities.
4.4. Dynamic Priority-based Task Selection Scheme
In DEMiR-CF, the robots make instantaneous decisions (from their local
perspectives) which are both precedence and resource feasible in the context
of the global time extended view of the problem. While the completion of the
mission is the highest priority objective, performance related objectives can
additionally be targeted. Each robot initially forms a rough schedule of its
activities for an overall time extended resolution of the mission. Since these
schedules are highly probable to change in dynamic environments and robots alsohave the real-time burdens of path planning, mapping etc., the formed rough
schedules are tentative and constructed by computationally cheap methods
(explained in Section 4.4.1.). Therefore, the robots in our framework come up
with their rough schedules and refine their plans during actual fast execution
when information available in the current context enables them to make specific,
detailed decisions.
Task selection and allocation is performed by evaluating each task according tothe selected cost function. Depending on the objective, different cost functions
can be defined, from the simplest functions to more complex and composite
evaluations. Cost evaluation is one of the key issues to make the framework
suitable for different domains in an efficient and effective manner. Cost functions
are analyzed in Section 4.5.
Since schedules are subject to change, we propose an approach in which tasks
are not scheduled initially but instead allocated to robots incrementally, without
ignoring the overall global solution quality. Therefore, the main objective
becomes determining a particular task to be assigned whenever it is convenient
in a precedence and resource feasible manner, instead of scheduling all the tasks
from scratch. Although not a concern during assignments, preemption (i.e.,
yielding) is possible to maintain the solution quality and to handle failures during
execution. Therefore, the allocation problem turns into a selection problem andis stated as the CTSP , which is introduced in Section 2.
Note that the CTSP presented earlier is an optimization problem as in
ScP , and it is desirable to find a solution by considering the problem from
the global perspective. Therefore, the instantaneous task selection scheme
needs to be strengthened by considering the problem as a whole, with the
designed cost evaluation functions. Depending on the objective function, either
priorities or penalties can be applied to find a near-optimal solution ensuring atime-extended view of the problem. This issue is analyzed in detail in Section 4.5.
The following definitions are needed to present our formulation to solve the
CMAP .
Definition (suitable task and suitable robot ) ti is a suitable task for robot r j , if
reqcapi ⊆ cap j and r j is a suitable robot for ti.
Definition (executable task ) ti is an executable task, if at least reqnoi number
of robots can be assigned for its execution.
Definition (task in execution ) tiej is a task in execution by robot r j or coalition
C j. T ie is a union of tasks in execution.
Definition (eligible task ) tEj is an eligible task, if it is an executable task and is
neither in execution (tie) nor achieved. T Ej is a union of eligible tasks for robot r j.
Definition (ineligible task ) tφ is an ineligible task if it is not an executable task , if
it is already achieved or if it is not a suitable task . T φ is a union of ineligible tasks.
Definition (predecessor task set ) P (ti) is defined as the set of all predecessor
tasks of task ti.
Definition (active task ) tAj is an active task if it is suitable, executable and
tasks in P (tAj) are completed. T Aj (⊆ T Ej ) is a union of the active tasks forrobot r j.
Definition (inactive task set ) is the set of all all inactive tasks (tIj ),
T Ij = T Ej \ T Aj contains the tasks that are suitable but not executable yet for
robot r j.
Definition (critical task ) A critical task tCj is a task that has inflexibility from
the point of view of resources and robot r j is suitable for that task. T Cj is a
union of critical tasks for robot r j.
Definition (rough schedule) A rough schedule S Rj for robot r j is a priority
queue of mission tasks that r j assumes it will execute.
4.4.1. Rough Schedule Generation Scheme
Each robot r j generates its rough schedule as a dynamic priority queue similar to
runqueues, by considering its critical task set (T Cj ), the eligible tasks (T Ej ), the
conjunctive arcs (if any) and the requirements. If there are no new online tasks
or invalidations, the order of the tasks which are connected by the conjunctive
arcs remains the same in the priority queue, even though there may be additional
intermediate entries into the queue at runtime.
Since each robot r j has different capabilities, the eligible task sets (T Ej ) and
the priority queue entries will be different. Sometimes uncertain information
(e.g., related to a local online task) or unexpected (internal or external) events
(e.g., detection of a fuel leakage) may result in this difference, even when robots
possess the same type of capabilities. The critical tasks may be determined either
by negotiations or by beliefs. Critical task information is used for determining
the task requirements such as power, fuel etc.
Intuitively, robots do not deal with the ineligible tasks (T φ), the union of tasks
that are already achieved or those that are not eligible from the capabilities
perspective while forming the rough schedules. The eligible tasks (T Ej = T \ T φ)
for robot r j consists of active and inactive tasks.
The rough schedule of a robot is generated by execution of Algorithm 1. curcs j
represents the remaining capacity of robot r j and reqcs(ti) represents the required
capacity for task ti in terms of the consumable resources (e.g., fuel). In the rough
schedule generation algorithm, while the rough schedule is being formed, theremaining capacity of the robot is also monitored. If the capacity of the robot is
input: Eligible task set (T Ej ), active task set (T Aj ), critical task list (LCj ), remainingcapacity (curcs j) of robot r j
output: Rough schedule (S Rj) of tasks, the top most suitable active task ts
ts = φ; R = curcs j; achievable = true;S Rj = GeneratePriorityList(T Ej , T Aj)/*Determines if the mission is achievable*/for each ti ∈ LCj do
R = R − reqcs(ti)if R < 0 then
achievable = false
R = curcs j
breakend if
end for
if S Rj = φ and (top(S Rj) ∈ LCj R − reqcs(top(S Rj)) ≥ 0) thents = top(S Rj)
end if
not sufficient for executing all of its critical tasks and the mission is believed to
be unachievable accordingly, then the robot may select an active task to execute
even if it is not a critical task for the robot in case new robots can be deployed.
However, if the mission is believed to be achievable, the robot may select to stay
idle until its critical tasks become active. This selection is done after forming
the rough schedule. The active task on top of the rough schedule that can beexecutable is the most suitable task to be executed for the robot. Sometimes the
rough schedule of the robot may be empty. In this case, the robot selects to stay
idle as determined in the DPTSS algorithm. The priority values to form rough
schedules are determined based on the mission and the objective function which
will be explained in Section 4.5.
4.4.2. DPTSS Algorithm
In our incremental allocation approach, the fundamental decision that each robot
must make is the selection of the most suitable task from the active task set ( T A)
by considering eligible task set (T E ). Algorithm 2 presents the DPTSS in which a
rough schedule is generated before making a decision. The four different decisions
made by robots after performing DPTSS are:
• continue to execute the current task (if any),
• join a coalition,
• form a new coalition to perform an available task, or• stay idle.
offer solutions for instantaneous assignments. DEMiR-CF can address both types
of classes by implementing incremental allocation of tasks with efficient bidding
strategies. Therefore, global solution quality is achieved from a time-extended
view of the problem by means of bid considerations. However, the approach is
also capable of offering solutions for instantaneous changes on task descriptions.Therefore, our framework is classified as capable of addressing both classes.
Incremental assignments eliminate redundant considerations for environments
in which a current best solution is highly probable to change, and efficient
and intelligent bidding strategies ensure solutions to be close to optimal with a
time-extended view of the problem in a computationally tractable way.
It is shown that by an efficient bid evaluation approach, globally near optimalsolutions can be observed in an auction-based approach. This is validated in the
following chapters by evaluating the performance of the framework in different
domains.
4.5.1. Cost Function Design Criteria
Depending on the selected application domain and a regular objective function,
cost functions should be designed appropriately. Different types of rough schedule
generation schemes can be performed by using the designed cost functions. At
this point, it is necessary to distinguish the rough schedule generation schemes
for independent tasks and interrelated tasks with workforce constraints.
4.5.1.1. Independent Tasks with Combinatorial Structures
There are several Operation Research methods to allocate independent tasks to
robots.
Integer Programming Formulation . Optimal results can be obtained
by an efficient Integer Programming (IP) formulation in IP solvers (e.g., the
commercial IP solver CPLEX (ILOG-CPLEX-9.0-UserMan)).
Branch and Bound Algorithms. In these algorithms, a search tree is
enumerated (branching) by constructing the smaller subtrees (subproblems)
within a feasible search space and searched through investigating lower and
upper bounds of each subtree (bounding). The procedure continues until all
nodes are either pruned or solved. The performance of the approach is dependent
on the branching and bounding algorithms (Toth and Vigo, 2001).
Heuristics . There are several heuristic approaches to find approximate solutions
to the optimal solution. The methods are classified into two subclasses: Classical
Heuristics and Meta-heuristics. In the classical heuristic approach, standard
construction and improvement methods are applied to the solution. These
approaches perform limited exploration of the search space and typically producegood results within modest computing times. In the meta-heuristics approach,
the algorithm searches a large solution space. Evolutionary algorithms, Tabu
Search and Simulated Annealing methods are the common methods belonging to
this class. Although the quality of the solution is much higher than that of the
classical approaches, time complexity increases dramatically in these approaches.
The applied procedures are usually context dependent and require finely tuned
parameters (Toth and Vigo, 2001).
All these cost evaluation procedures can be used in DEMiR-CF. Each robot
may simply perform the explained operations to generate the rough schedules
and select a task to perform. However, as experiments in the following chapters
illustrate, classical heuristic cost evaluations, which are more applicable to
robots, produce highly acceptable and efficient results.
4.5.1.2. Interrelated Tasks with Multi-Robot Requirements
The interrelations among tasks and resource requirements are represented as
directed acyclic graphs in each robot’s world knowledge. The generated rough
schedules respect the precedence and resource constraints. For the makespan
objective in general, branch and bound methods can be applied. However, since
there are interrelations and resource dependencies, reaching a consensus by
communication may sometimes be intractable.
There are two efficient heuristic methods to generate feasible schedules:
Forward-Backward Schedule Generation Schemes (Pinedo, 2005) for the RCPSPand another method that we propose which generates a topological sort of
the directed acyclic graph fed by different priority rules. There are various
priority rules that can be applied (Brucker and Knust, 2006) for evaluating cost
values to select tasks. These are classified into four types: activity-based rules,
network-based rules, critical path-based rules and resource-based rules. Different
priority rules are listed as follows: activity-based rules by selection of tasks
with the smallest processing time (SPT) or the longest processing time (LPT);
network-based rules by selection of tasks with the most immediate successors(MIS), the least immediate successors (LIS), the most total successors (MTS),
the least total successors (LTS) or the greatest rank positional weight (GRPW)
(the largest processing time of all successors); critical path-based rules with
the smallest earliest starting time (EST), the smallest earliest completion time
(ECT), the smallest latest starting time (LST), the smallest latest completion
time (LCT) or minimum slack (MSLK); resource-based rules with the greatestresource requirements (GRR).
4.6. Task Allocation
DEMiR-CF uses the standard auction procedures of CNP (Smith, 1980) to
announce the intentions of robots on task execution and to select the reqno
number of robots for a coalition in a cost-profitable, scalable and tractable
way. Additionally, Plan B Precaution Routines are added to check validity,
consistency and coherence in these negotiation steps. Each robot intending to
execute a task announces an auction after determining its rough schedule and
performing the DPTSS.
4.6.1. Distributed Task Allocation Scheme
Basically, auction announcements are ways to illustrate intentions to execute
tasks for which reqno = 1 or to select members of coalitions to execute tasks
for which reqno > 1. Therefore, if more than one robot declares intentions to
execute the same task, the more suitable one(s) is selected in the auction by
considering the cost values. Auction negotiations and selection of the suitable
robots are performed in a completely distributed manner by the auctioneers.
Single task items are auctioned and allocated in auctions. Auction negotiation
implemented in the framework consists of standard steps to clear an auction.
Robots can get the necessary task details from the auction offers, and then
check the validity of the auction. If the auction is invalid, related precaution
routines (explained in Section 4.8.) are activated. Otherwise, the candidaterobots send their cost values as bids. (The other candidate robots may behave
as auctioneers as well. If the auctioneer does not receive the required number
(reqno) of bids (also counting in its own bid) from the other robots by the
predefined deadline, it cancels the auction. Otherwise, it ranks all bids and
assigns the best suitable robot with the lowest cost value to the executable
task (if reqno = 1), or suitable coalition members (if reqno > 1). The
framework allows multiple auctions to be carried out simultaneously. The basic
steps of an auction negotiation process are illustrated in Figure 4.4. Validitycontrols are performed to ensure system consistency as a part of the Plan B
Figure 4.5: An invalid auction for an achieved task is canceled by the auctioneerafter being warned by one of the candidates
formats.
Most of the time, the broadcast type of message propagation is used in
DEMiR-CF. For some special negotiations peer-to-peer messaging is also used.
The message types designed for DEMiR-CF are given in Table 4.3.
4.6.3. Roles
Members of coalitions are selected by auctions. The auctioneers are also active
robots in the system. A robot (r j) may take different roles for task ti, such as
auctioneer, bidder (Bij), coalition leader (CLi) or coalition member (CM i) duringruntime.
• An Auctioneer is responsible for managing auction negotiation steps and
selecting reqnoi of suitable members of a coalition.
• A Bidder is a candidate to become a member of a coalition to execute a
task.
• A Coalition Leader is the robot responsible for maintaining the coalition
and providing synchronization while performing the coalition task.
• A Coalition Member is one of the members of the coalition, and it executes
a portion of the task.
A robot r j may be in more than one Bij roles for different tasks, but may not
be in multiple roles as an auctioneer, a CM i or a CLi at the same time. The
auctioneer is responsible to select the required number of robots (the coalition
leader and members) for task execution. The auctioneer may or may not take
place in the coalition for which it offers an auction. The coalition leader, selectedby the auctioneer as the one with the minimum cost for executing the task,
maintains the coalition and keeps track of the members’ conditions and their
updated information. After the execution of the task is completed, the coalition
ends. Each robot is allowed to take part in only one coalition until it leaves
the coalition. Coordination between coalition members is implemented through
synchronization messages. We assume that robots are not able to infer the stateof others by observation, although such capabilities would only provide more
In the framework, instead of using complicated re-allocation procedures,
we propose incremental selection and task switching schemes for behaving
myopically while thinking globally using bid evaluation heuristics. Provided
with an efficient bid evaluation heuristic, the dynamic task selection scheme
ensures task switching whenever it is profitable. Each robot, independent of its
current status from executing a task or not, can offer a new auction or select to
execute a task already being executed by another robot with a worse cost value
than it will cost for itself. If task switching occurs with a coalition member, the
corresponding coalition member is released from the coalition and becomes a
suitable robot for other tasks.
To ensure maintaining the solution quality against environmental changes, wepropose a dynamic reconfiguration approach. Coalition members can leave a
coalition when there is a sufficient number of members to execute the task. The
coalition leader is responsible for broadcasting the maximum cost of execution for
the task by one of the coalition members in each execution step. In the decision
stage, each robot receiving these messages evaluates the maximum cost value of
each coalition. If a robot detects that its cost is lower than the maximum cost of
the coalition and it is released from its current coalition, it sends a join request
message to the coalition leader. The leader, getting a join request message,directly adds the robot to the coalition. If the coalition leader detects that the
size of the coalition is larger than required, it can release coalition members with
the maximum cost value for the current task. Getting a released message, a robot
can proceed to select another suitable task. When the coalition leader considers
the size of the current coalition, it also checks the failures. Since each robot in the
coalition broadcasts an executing message along with the updated cost values,
their failure can be detected by the coalition leader. The failed robots are also
released from the coalition. If the number of members to execute the task is
below the required number for a period of time, the coalition leader cancels the
coalition. An illustrative example of coalition reconfiguration is given in Figure
4.6. Such a situation may occur if a robot is not reachable when the auction
announcement is made (a). When the situation changes (b), the robot may take
over the role of the member with the maximum cost value in the coalition (c).
The released member can select another task to execute.
Robot with the maximum cost value (maxcost1) in the coalition
Figure 4.6: An illustrative reconfiguration scenario for maintaining thecoalitional solution quality
The decision regarding which robot/agent is a member of which coalition
(task execution) is an important issue. Since coalitions are disjoint in our
case, assigning a robot/agent to a coalition may prevent another advantageous
situation in which one of the already assigned robots may take a role in a near
future formation. Further negotiations, other than auctions are needed to reachconsistent agreements.
One important issue that should be addressed in robot systems is ensuring ways
to plan for the global problem from local views. From our point of view, this
can be achieved through extensive bid evaluation designs and additional routines
to improve solution quality. We generalize statements given in desJardins et
al. (1999) for continual planning. Added to the integration of planning and
execution, task allocation/scheduling should also be integrated into a continualplanning process. Currently most multi-agent/robot coordination architectures
implement decomposition, allocation and execution steps sequentially, and
respond to the contingencies (mostly embedded in the model) in the execution
phase. However, this integration is provided in DEMiR-CF by means of the Plan
B Precaution Routines which are explained in the following section.
4.8. Plan B Precautions: A System-wide Contingency Handling
Mechanism
Plan B Precautions are taken in DEMiR-CF by The Model Update Module which
updates the system model of the robot and The System Consistency Checking
Module. The Model Update Module uses incoming information and its own
perception data to update the models of the robot. The System Consistency
Checking Module provides warning messages to keep the system consistent.
In DEMiR-CF, information is not assumed to be complete and robots allocate
and execute tasks in a distributed manner. This framework can take advantage
of communication when it is reliable. The consistency of the current task states
is ensured by Plan B Precautions which are carried out by the robots in a
completely distributed manner. Consequently, the system immediately responds
to various failure modes and can recover from them. Current implementation
uses explicit communication to detect conflicts and contingencies. However,
failures in communication can also be handled by precaution routines. (If robotscan observe each other implicitly, model updates can be implemented in a
similar manner.) Appropriate precaution routines according to the contingent
situations are activated to either correct the models or to initiate recovery actions.
4.8.1. Representation of The System Model In Each Robot’s World
Knowledge
Each robot keeps the models of the system tasks and other robots in itsknowledge base. Models of different robots may become inconsistent because
of the uncertainties, incomplete knowledge, assumptions etc. It is not always
possible to share a common world knowledge in decentralized systems as in the
case of ours.
Task models are stored as Finite State Machines (FSM) where each task can be
in a different state. Task FSMs are illustrated in Figure 4.7 and Figure 4.8 for
single-robot and multi-robot task models, respectively. The difference in these
two types of representations is mainly the synchronization issue that robots needto achieve when participating in a coalition. The state transitions are activated
• self inexec: A task in this state is being executed by the corresponding
robot in a coalition and synchronized or the coalition involves only the
robot itself.
• others inexec: A task in this state is being executed by other robots.
• achieved : This is the final state of the tasks that are achieved.
• unachievable: This state is used for the tasks that are not traversable or
achievable.
unachievable
message
is received
self_auctionedawarded to
Itself in
a coalitionawarded
self_inexec_wait_
sync
available
threshold
Is exceeded
released
from
the coalition
selected
by the DPTSS
achieved
achieved messageis received
init/
auction
is canceled
others_inexec
valid execution
message
is received
coal execution is
canceled
uncertain
others_auctioned
self_inexec
unachievable
auction message
is received
awarded by
another robot
robot believes that
the task is unachievable
synchronization
is achieved
ready for
execution
task is
achieved
valid execution
message
is received
auction message
is received
selected
by the DPTSS
threshold
Is exceeded
awarded to a coalition
without itself
valid execution
message
is received
joined
to the
coalition
Figure 4.8: States of the FSMs for multi-robot task models
Robot models are also stored in FSMs, where each robot model has a state whichis assumed by the corresponding robot. A robot may be in one of the following
states:
• idle: A robot is assumed to be in this state when there is no evidence that
it is executing any task.
• executing : A robot is assumed to be running properly and executing a task
whenever there is supporting evidence.
• failed : A robot is assumed to have failed when there is no evidence that ithas been running properly for a long time.
5. EMPIRICAL EVALUATION OF DEMiR-CF ON THE MULTIPLE
TRAVELING ROBOT PROBLEM
Search and rescue operations, space exploration, and reconnaissance/surveillance
applications require effective multi-robot exploration. Although the coordination
problem seems to be similar, the overall objective for cost optimization may
be different in these domains. Search and rescue operations may require time
minimization, while space operations require minimization of total path length
traversed by all robots being proportional to the total energy consumed by robots.
In this chapter, we investigate the performance of several heuristic functions
integrated into the framework for multi-robot exploration tasks as a case study
(Sariel and Balch, 2005a, Sariel and Balch, 2006c). Because this problem area
is well studied in operations research, optimal solutions are available, so we can
analyze the deviations of results generated by the framework from optima. Note,
however, that optimal solutions sometimes require time consuming computations,
while our solutions can be found quickly. Although the problem domain consists
of the same types of tasks that can be executed by a homogeneous team of robots, still it is NP-Hard due to the combinatorial structure of the problem.
5.1. MTRP Problem Statement
The single robot exploration problem, a variation of the well known NP-Hard
Traveling Salesman Problem (TSP), is to find the minimum cost traversal of a
given number of targets (T ) without considering the return cost from the last
target to the initial location for a single robot. The problem can be stated as
finding the minimal Hamiltonian path on a given fully connected graph with all
nodes to be visited. The travel costs are assumed to be symmetric. Although
the TSP is NP-hard, there are many efficient k-OPT heuristic methods in the
literature (Lawler et al., 1985).
The Multi Traveling Robot Problem (MTRP) or the multi-robot multi-target
exploration problem is a generalization of the TSP in which there is a team
of robots (R) to visit targets (ti ∈ M MTRP ) at least once (ideally at mostonce). This problem may be stated for different objectives such as minimizing
Figure 5.2: Clustering the targets does not necessarily result in the optimalsolution
The optimal allocation can be obtained by clustering the targets. Clustering
methods can be used to form target clusters. However, these techniques use
distance information. Therefore, in some situations, clustering methods also do
not produce optimal results because of ignoring additional costs as in Figure 5.2.
Lagoudakis et al. (2005) present an extensive analysis of the multi-robot
exploration problem from the point of view of solution guarantees. In their work,
they analyze allocation approaches for both sequential tree construction androute generation, and direct allocation while constructing paths. From our point
of view, as it is suggested in their work, generating a Minimum Spanning Forest
(MSF) and constructing routes on separate MSTs may not always be an effective
method. One reason is that there may be different MSF solutions for the MTRP
case in which some of the distances are the same. It may result in different
allocation alternatives, and if there is not a reasonable allocation strategy
other than selecting the minimum distance, the solution quality may degrade
accordingly. The other reason is that while allocating targets, considering only
distance is not an effective approach because additional costs added in the route
construction phase are ignored. A sample situation is given in Figure 5.3.
Even though constructions of the tree-like structures are computationally
efficient, tree construction and making allocations from scratch may result in
suboptimal allocations. An IP approach may be used for finding optimal
allocations. However, for even moderate sized instances, there is no guarantee of
the solution time. Changes in the distance values between target and robot pairs
may be frequent in dynamic environments. In this case, the solution should bereexamined. Even very small changes may completely change the overall solution.
In this formulation, R denotes the set of robot vertices and T the set of target
vertices. xij is an indicator (0/1) variable for i ∈ T ∪ R and j ∈ T . If xij = 1,
then location j must be visited directly after location i. cij is the cost value to
traverse between i ∈ R ∪ T and j ∈ R ∪ T .
The first set of constraints ensures that target locations are visited exactly once,
while the second set ensures that robot and target locations are left at most once
and, finally, the third set guarantees that there are no cycles among the target
locations (subtour elimination constraints) (Lagoudakis et al., 2005).
5.1.2.2. Branch and Bound Approach
The MTRP problem can be solved by Branch and Bound algorithms. In these
algorithms, a search tree is enumerated (branching) by constructing the smallersubtrees (subproblems) within a feasible search space and searched through
investigating lower and upper bounds of the current subtree (bounding). The
procedure continues until all nodes are either pruned or solved. The performance
of the approach is dependent on the branching and bounding algorithms (Toth
and Vigo, 2001).
5.1.2.3. Heuristics
There are several heuristic approaches to find approximate solutions close to
the optimal solution. The methods are classified into two subclasses: Classical
Heuristics and Meta-heuristics. In the classical heuristic approach, standard
construction and improvement methods are applied to the solution. These
approaches perform limited exploration of the search space and typically produce
good results within modest computing times. In the meta-heuristics approach,
the algorithm searches a large solution space. Evolutionary algorithms, Tabu
Search and Simulated Annealing methods are the common methods belonging to
this class. Although the quality of the solution is much higher than that of the
classical approaches, time complexity worsens dramatically in these approaches.The applied procedures are usually context dependent and require finely tuned
parameters (Toth and Vigo, 2001).
5.1.3. Robotic Research Methods for the MTRP
Operations Research methods are applied and integrated into robot systems in
The Prim Allocation method (Lagoudakis et al., 2004), based on Prim’s
Algorithm (Jarnik, 1930; Prim, 1957), generates an MSF of targets and robots.
An MSF consists of separate robot trees. These trees are constructed by adding
each unallocated target to the closest robot path containing the node with the
minimum distance to the target, until all targets are allocated. In other words,
a new target is added by considering the distances between the target and the
nodes of the robot tree instead of considering the last position of the robot.
Each robot offers an auction for a target and one of the targets is allocated in
each round. Before robots run and visit the targets, all targets are allocated.
Whenever the world knowledge is changed, the remaining unvisited targets
are reallocated using the same algorithm. Like Prim’s Algorithm, the Prim
Allocation method is bounded by 2*OPT for the MTRP. Since this methodoffers a single item allocation approach, it is the best candidate in the literature
to compare the task allocation approach of DEMiR-CF. Furthermore, the Prim
Allocation approach is discussed with its performance bounds and the details
necessary to implement it, so we have been able to implement and compare the
performance of DEMiR-CF to that of the Prim Allocation method.
The depth-first traversal solution of an MST is bounded by 2*OPT, and the
traversal and subtree selection does not affect the solution quality in solving theTSP. However, for the open loop version of the TSP, as in the MTRP, selection
of the subtree that is traversed affects the solution quality to a great extent.
To improve the solutions, traversal may begin with the shortest depth subtree
and continue with traversal on the subtree with the next shortest depth. A
sample situation is given in Figure 5.4. We added this improvement to the PRIM
Allocation method while making the comparisons.
Figure 5.4: Effects of the MST traversal strategy on the total cost for the opentraversal version of the TSP
Both combinatorial (Dias and Stentz, 2002, Berhault et al., 2003) and single
auction methods are studied for the MTRP in the literature. In GRAMMPS
(Brumitt and Stentz, 1998), one of the earliest works to solve the MTRP, a mission
planner works centrally either on one of the robots or as an operator. The mission
planner selects a robot for each target. The system can regenerate plans when the
environment changes. Authors claim that for the initial state of the system, the
allocations may be suboptimal. However, in later steps, when the number of open
missions decreases, the system can find close to optimal solutions. By using the
simulated annealing algorithm, a randomized search over all possible allocations
is made. In their latest work, Dias and Stentz (2002) propose a market-based
scheme introducing a leader approach for combinatorial task exchanges. These
leaders are responsible for multi-party multi-task optimizations for obtainingoptimal results. In combinatorial auctions, different combinations of tasks are
offered and bidders bid by considering all tasks in these combinations. Thus, this
method may become intractable for large instances or for dynamic situations in
which calculations should be made frequently. Especially when the environment
is dynamic, allocations may become suboptimal. Then, a combinatorial exchange
mechanism is necessary to maintain optimality. Computational requirements for
combinatorial auctions increase dramatically for dynamic environments.
5.1.4. Cooperation Objectives
Different types of objectives complementary to the main goal may be selected
to optimize the performance of the system, as in scheduling problems.
Examples of such objectives are total energy minimization (total path length
minimization), time minimization, average energy minimization, makespan
minimization (Lemaire et al., 2004), etc. Based on the selected objective function,
cost evaluation may need to be designed differently. In Figure 5.5, the paths that
robots traverse in order to optimize target allocations for two different objectivesare shown. The robots are located at the bottom of the figure in the initial
configuration facing the three targets to be visited. Achievement of the total
path length minimization objective is illustrated in the first row of the figures
(a-e), whereas in the second row of the figures, robots minimize the time needed
to achieve the mission (f-j). When the total path length is minimized, one of
the robots visits all targets. For the time minimization, all robots are involved
in the target visiting process. Related videos of these runs are available at the
Figure 5.5: Two different optimization objectives for the MTRP: The first row of the figures illustrates achievement of the total path length minimization objective.In the second row, robots minimize the time needed to achieve the mission.
In a more general setting, incorporating composite objectives (sometimes for
pareto optimal solutions) into the system may be more useful for the success
of the mission (e.g., both the path length and the time need to be optimized).
Besides these objectives, there are other real-time execution issues that should
also be analyzed. Some of them are listed below:
• reducing collision risk,
• target priorities/strategic target selection,
• time windows, and
• multi-robot requirements to visit a target, etc.
5.1.5. Application Domains for the MTRP
The MTRP domain forms an important basis for many other domains
such as Search and Rescue (SR), Space Explorations, Object Construction,
Pick-up/Delivery, etc. ”Goto Target” task is one of the main tasks for these
domains. Many high level tasks or assemblages of behaviors require this task to
be performed. Even “pushing an object” task can be performed by simply going
to the next point in the desired pushing path; that way the object is assumed to
be pushed when the target location is reached.
In SR operations (Davids, 2002), the mission is challenging due to short time
constraints and success is crucial. Victim locations in a disaster area are not
known in advance. Incoming sensor data can be used to determine probabilistic
information regarding the possible victim locations (Sariel and Akin, 2005).
In practical applications, computing the true optimal solutions is not always
required due to several reasons (Reinelt, 1994). Those reasons may be the
incorrect modeling of the underlying problem (targets) or the lack of sufficient
time to find the optimal solution. These are common cases in robot applications
along with the real-time issues presented in Chapter 2. To meet all these
limitations, we propose a dynamic and distributed task allocation scheme,
DEMiR-CF, to coordinate robots that cooperate to fulfill different parts of a
mission. Dynamism is achieved through the incremental selection and allocation
of the targets. By means of the distributed characteristic of DEMiR-CF, each
robot is allowed to select a candidate task for itself and, next, the robots proceed
to cooperate in the process of selecting the most suitable robots for the tasks.
DEMiR-CF is designed with the capability to deal with the situations presented
in Chapter 2. The framework can efficiently respond to these events and the
solution quality is maintained simultaneously with the real-time task execution.
We propose a general mechanism for multi-robot cooperation for the MTRP but
not necessarily specific heuristic functions to solve the problem, although we
validate their successes.
5.2.1. The Dynamic Priority Based Task Selection Scheme
The dynamic priority based task selection and allocation scheme ensures two
types of selection. Each robot selects a target to visit. The system provides
a CNP-based selection method to select the appropriate robot to allocate a
task. This mechanism ensures distributed, robust and scalable allocations. The
incremental assignment approach eliminates the complexity of the decision of
allocating all targets to all robots at a time.
5.2.1.1. Selecting is Eliminating the Other: Incremental Allocationthrough Unconditional Commitments
Each robot selects candidate targets that are suitable for itself and forms a target
set (rough schedule). This set consists of the targets suitable to be visited by the
robot. The robot then selects the most suitable candidate task to perform among
these targets in its corresponding set. After selecting the most suitable target for
itself, the robot announces its intention. A CNP-based selection is implemented
and the most appropriate robot among the team of robots to perform the task
(to visit the target) is selected. When robots receive task execution intentionmessages, they either send their cost values as bids for the announced target or
only considers the distance between targets in T Rj and the robot r j.
c ji = dist(r j , ti) ti ∈ T Rj (5.3)
The FAC (Farthest Addition Cost) heuristic function considers costs as if there
is a route for T Rj as in Equation 5.4 and applies a penalty for not visiting the
boundary targets. Boundary targets, tb1 and tb2, are the targets in T Rj with the
maximum distance value. The FAC heuristic forwards robots to these targets in
T Rj to some degree. The main idea behind this approach is that the open loop
traversal should contain both tb1 and tb2s. If the robot heads towards one of these
targets, if profitable (α), the longest path can be traversed by traversing other
targets on the path. A sample illustration of this cost function is given in Figure
5.6. In this figure, although t2 is closer to r1 than t1, with the FAC heuristic
applied, t1’s cost value is smaller than that of t2 (3 < 3.6), hence resulting in
a better route shown by the dashed arrows. We have performed an empirical
analysis of the parameter α; the best results are observed for a value of 0.6.
c ji = α ∗ dist(r j, ti) + (1 − α) ∗ [dist(tb1, tb2)
−max(dist(ti, tb1), dist(ti, tb2)] (5.4)
{dist(tb1, tb2) = max(dist(tk, tl)) ti,k,l ∈ T Rj}
2 3 4
4
3
5
(T )R1
t m1 tm2
{c = 5*0.6+0 = 3}
{c = 4 * 0.6 + (12 - 9) *0.4 = 3.6 }
r1
11
12
t2t =1t3 t4 = t5
Figure 5.6: Target selection strategy by the FAC Heuristic function
5.2.1.3. Dynamic Task Selection and Distributed Target Allocation
Each robot executes Algorithm 3 to generate its rough schedule. The robot then
selects the most suitable candidate task (ts, the most suitable target among the
rough schedule targets) to perform.
Algorithm 4 forms the main loop for incremental task allocation and it is calledin the beginning of the mission execution and whenever the world knowledge
After selecting the most suitable target for itself, each robot announces its
intentions by a single-item auction. Selection of the best robot for a task is
performed by using the Contract Net Protocol (CNP) in our approach. Although
the CNP presents the formalism on the relationships between managers and
contractors, some simple decisions are left to the designer. Most auction basedtask allocation schemes offer solutions for allocating one/a subset of task(s) of
the overall mission. However, there is usually little information about when task
announcements and reassignments are made.
The approach we propose allows for multiple auctioneers and winners for different
tasks, depending on the optimization objective. In the case of the total path
length minimization objective, ending one auction at a time results in better
performance when there are relations between targets. This is the reason whysome auctions are canceled when there are multiple auctions going on at the
same time. On the other hand, auctions/executions are canceled only if there are
relations between the targets in consideration for the time minimization objective.
These two different responses are generated by Algorithm 5 and Algorithm 6,
respectively.
Algorithm 5 Response for Path Optimization - r j
if auction/execution is in progress & (c ji > clk) thencancel auction/execution for ti
end if if c jk < clk then
send bid value for tk
end if
Algorithm 6 Response for Time Optimization - r j
if auction/execution is in progress & (c ji > c jk ) || ((c ji > clk) & (tk or rl is
close to the T Rj)) thencancel auction/execution for ti
end if if c jk < clk then
send bid value for tk
end if
5.2.2. Failure Detection and Recovery
Plan B Precaution Routines integrated into the dynamic task selection mechanismis applied for the MTRP as explained in Chapter 4. The single-robot task models
are used to represent the states of the FSMs for individual tasks. Robots
continually perform model updates and consistency checking operations using
information received from the incoming messages and the data perceived from
the environment during runtime. When robots receive task execution intention
messages as auctions, they either send their cost values as bids for the announcedtargets or send warning messages to the sender robots. Failures can be detected
immediately and recovery routines are activated for the failures or contingencies.
5.3. Bounds on the Solution Quality
The performance of the Prim Allocation algorithm is proved to be bounded by
2*OPT (Lagoudakis et al., 2004). The difference between the Prim allocation and
the CC heuristic approach is in the robot location considered. Our framework
combined with the CC heuristic considers the latest location of robots, while
the Prim Allocation algorithm considers their initial locations. Assuming there
are two subtrees of the MST, the CC heuristic approach forwards robots into
either of the subtrees of the MST, leaving the other subtree to be traversed by
another robot or by itself. If in the end, the first robot traverses the subtree,
the solution cost is the same as the Prim Allocation solution cost. However, if
another robot traverses the other subtree, the generated solution is better than
depth-first traversal since the other robot has been favored because of its cheaper
cost value. The CC heuristic can be classified as one of the BidSumPath heuristics(Lagoudakis et al., 2005) and it is shown that the solution is bounded by 2*OPT.
The FAC heuristic forwards robots to one of the border subtrees. In the worst
case scenario, the next selection phase forwards a robot to the next subtree in
the MST before the completion of the traversal of a subtree (usually this results
in the elimination of long connections among subtrees and gives better results).
However, by making use of triangle inequality, going back to the previous subtree
cannot be greater than two times the traversal of the corresponding MST edges
in this worst case.
5.3.1. Analysis of the Approach
The approach we have proposed for the MTRP offers a polynomial time solution.
Sorting the distance values to find the boundary targets takes O(nlog(n)) for all
n number of tasks. Both cost and queue initialization are implemented in O(n).
Top element selection and deletion is performed in O(logn). Therefore, the total
complexity is bounded by O(nlog(n)). In the worst case, the environment is
dynamic and cost values change frequently in the order of O(l). Then the totalcomplexity becomes O(lnlog(n)) for each robot.
“Morse decomposition”and spinning behavior is stated to be efficient for searching
an area (Acar et al., 2002). This is used in our system for locating objects whose
estimated locations are not exactly known or when there is uncertainty about the
location of either an object or a robot. This idea is useful for locating specific
objects that are very likely to be located in the search area. The framework
implements the Archimedean spiral in robot behavior, an is inspired from the
geometric implementations given in Figure 5.7.
Figure 5.7: The Archimedean spiral and two simple implementations
A search and rescue domain in which victim locations are not exactly known but
may be estimated, is a candidate domain for applying this idea. However, if therobot is only visiting waypoints, spinning behavior may be time consuming in the
application.
Figure 5.8: The robot can directly find the target with a reasonably accurateestimation
We define a deadline threshold for the spinning behavior in our implementation.
The robot searches the area to find the object until the deadline is reached. If
the object is not found within the deadline, the robot gives up its search for that
target and continues to look for other targets, if any. In Figure 5.8, the robot
nodes; the number of robots change in the range [1-50] and the number of targets
in [10-50]. Time comparison results are taken for the centralized implementation
of the methods. The distance calculations among targets and robots are excluded
from the run time period while the IP model generation is included in the time
period because it is part of the solution. Both approaches are assumed to berunning on the given distance matrices. The results are presented as averages of
100 independent runs for the randomly generated test instances. The running
time results are averaged over 30 runs.
In the third set of experiments, the real-time performance of DEMiR-CF on
the MTRP is evaluated as dynamic simulation experiments on the professional
mobile robot simulation software, WEBOTS (Michel, 1998). WEBOTS contains
a rapid prototyping tool allowing the user to create 3D virtual worlds withphysics properties, such as mass repartition, joints, friction coefficients, etc. The
fourth set of experiments is performed on real robots, namely Khepera II.
Table 5.1: FAC heuristic function performance results for the known TSPinstances
TSP Instance FAC Optimum% Deviation fromthe Optimum
ATT48 - 48 nodes 33537.83 31470.4 6.5
EIL51 - 51 nodes 444.01 413.51 7.37
BERLIN52 - 52 nodes 8104.99 7305.38 10.94
EIL101 - 101 nodes 725.31 629.38 15.24
FAC Heuristic Optimum
Figure 5.10: Open-loop routes of the FAC Heuristic function and the Optimum
Large standard deviation values for the IP approach present the dependency
of the solution time on the problem instance structure. The performance of
the IP approach is close to the worst case performance for some instances,
not given in these statistics. Depending on the application and the instance
size or the frequency of the changes in the distance values, the IP approachmay be impractical without guarantees on the solution time. Allocations by
using a heuristic approach (either DEMiR-CF or the Prim Allocation) can be
implemented in a very short time as expected and are given in Figure 5.11. This
graph presents running time results of the approaches for the single robot case.
With a decreasing amount of target densities, the IP approach solution time
decreases accordingly.
5025 15 5 1
5040302010
0
10
20
30
40
# of Robots
PRIM ORG
# of Targets
D e v i a t i o n f r o m O
P T
5025 15 5 1
5040302010
0
10
20
30
40
# of Robots
PRIM SD
# of Targets
D e v i a t i o n f r o m O
P T
5025 15 5 1
5040302010
0
10
20
30
40
# of Robots
DEMIR−CF + FAC
# of Targets
D e v i a t i o n f r o m O
P T
5025 15 5 1
5040302010
0
10
20
30
40
# of Robots
DEMIR−CF + CC
# of Targets
D e v i a t i o n f r o m O
P T
Figure 5.12: Performance results of the heuristic approaches
The overall performance results are given in Figure 5.12 as deviations from the
optima with standard deviation, averaged over 100 runs. PRIM-ORG values
represent the results of the Prim Allocation method without considering subtree
sizes on the traversal, while PRIM-SD values represent the results with shortestsubtree selection improvement. Results of the FAC heuristic approach are
used. The figure on the left is the total number of messages sent by the
robots. The surprising result here is the increase in the number of messages
for the NO PREC case. Although the contingency handling mechanism seems
to inject additional messages into the system at first glance (for ensuring the
system consistency), these results reveal that it also eliminates the redundancyin the number of messages for multiple bids and auctions. The figure on the
right shows the total path length traversed by the robots. The path length is
presented by discretized values (each unit is 70mm). The experiments reveal
that using the contingency handling mechanisms reduce both the number of
messages sent and the total path length traversed by the robots. This result
validates the efficiency of integrating the contingency handling mechanism into
the system. Dynamic simulation experiments where the robots perform both
0 2 4 6 8 10 120
100
200
300
400
500
600
700
800
900
1000
Number of Robots
N u m b e r o f M e s s a g e s
NO_PREC
USE_PREC
0 2 4 6 8 10 12100
120
140
160
180
200
220
240
260
280
300
Number of Robots
T o t a l p a t h l e n g t h t r a v e r s e d b y r o b o t s
NO_PREC
USE_PREC
Figure 5.15: Performance results for the Contingency Handling Mechanism
task allocation and real-time sensing, localization and mapping can demonstrate
how the knowledge of map information changes the problem instance and the
requirements on reallocations when new information is encountered by robots.
Two sample scenarios with 3 robots and 10 targets present this hypothesis. The
initial configuration of the environment is illustrated in Figure 5.16. The targetsare presented by the blue landmarks. Two robots are located at the bottom
of the environment and the third robot is located in the center of the environment.
In scenario 1, lacking the map information, the robots initially assume that
the environment is empty, even though there are obstacles. Thus, visiting the
predetermined target locations takes much more time than expected. The final
state of the environment at the end of the mission is presented in Figure 5.17.
Robot paths are drawn by the simulator in this figure. The robots carry outlocalization and mapping operations simultaneously to visit targets efficiently.
The actual paths and the estimated paths traversed by the robots using odometry
information and kinematics calculations are illustrated on the left side of Figure
5.18. The estimated paths are represented by dotted lines. As can be seen from
the figures, due to the slipping and encoder errors, the estimated paths do not
completely overlap with the actual paths traversed. In such cases, target locationsmay sometimes be incorrectly interpreted. The errors on the traversed paths may
be reduced by using landmarks in the environment and adjusting the localization
error accordingly. The occupancy mapping approach (Moravec and Elfes, 1985)
is used in our experiments. The maps generated by the robots are illustrated
on the right side of Figure 5.18. Rough schedules are generated and robots visit
targets based on the up-to-date selections. During mission execution, one of
the robots detects that a target is not traversable after discovering the obstacles
surrounding the target. In scenario 2, the map of the environment is given to therobots (Figure 5.19). As expected, robots do not intend to visit untraversable
targets. Paths and maps of this scenario are illustrated in Figure 5.20.
Figure 5.16: The initial configuration of Scenario 1
Figure 5.17: The final configuration of Scenario 1
5.5.4. Experiment 4: Real-Time, Real-World and Dynamic
Environment Experiments
Each Khepera II robot is equipped with a 25MHz MC68331 micro-controller,
512K Flash and 512K RAM memories and 8 infra-red sensors with a limited
obstacle detection range as simulated in WEBOTS. Communication is achieved
through wireless links. Real Kheperas have standard radio turrets mounted on
them to communicate through the selected radio frequency. Emitters in the
simulator are configured to have a baud rate of 9600 and a buffer size of 1024B
as in the receiver modules.
Figure 5.21: The Khepera II robot visits six targets in the environment
Figure 5.22: Three Khepera II robots divide the labor of visiting the six targets
Each robot executes a multi-threaded controller software to achieve the
functionality of DEMiR-CF. Different modules on the task allocation layer are
integrated with the low-level Sensory Interface, Motor Interface, Motion Model
and Mapping modules in the multi-threading structure.
The overlaid video frames from real robot experiments that we have conductedare illustrated in Figure 5.21 - Figure 5.23. Figure 5.21 illustrates a path
traversed by one Khepera II robot visiting six targets in the environment. The
target locations are shown with blue flags and fixed in all real robot experiments.
When we have a multi-robot team, the mission completion time is reduced by
the robots’ division of labor. The robot paths are illustrated in Figure 5.22. The
multi-robot team can successfully handle robot failures as illustrated in Figure5.23. All robot positions at the time of the failure are marked with red filled
circles.
Depending on the initial location of the robot, the path constructed to
traverse the targets differentiate for the single robot case. This is illustrated
in Figure 5.24. In this figure, each column (a-c) represents the illustration of
an independent run with a different initial deployment location for the robot.
The continuous video frames are divided into episodes in which the images of the videos are overlapped and the final overlapped image is illustrated (e.g.,
a1,a2..a6) in the figure.
Figure 5.25 illustrates the scenario (Scenario 1) with three robots for the same
setting of the target set. Time is minimized by using a three-robot team. Robots
use DEMiR-CF to select and allocate tasks in a distributed manner.
(a) (b) (c)
(d) (e) (f)
Figure 5.25: Scenario 1: the multi-robot case without failure
The failure of the third robot and the rest of the run are illustrated in Figure 5.26
(Scenario 2). In this scenario, the third fails after visiting its assigned task. The
failure of the robot is forced by the human agent isolating the related robot. At
the time of the failure, the other robots are busy with their own target visitingtasks. The failure of the third robot does not block the execution. Since the
allocations are performed incrementally in DEMiR-CF, the target that is assigned
to the third robot in Scenario 1 is not allocated to the third robot in this situation
because it fails immediately after achieving its first task. After the failure, the
second robot selects this target as an available target among the other targets.
There is no recovery in this scenario, but redundant allocation procedures areeliminated by DEMiR-CF.
(a) (b) (c)
(d) (e) (f)
O
Figure 5.26: Scenario 2: the multi-robot case in which the third robot fails aftercompleting its task
(a) (b) (c)
(d) (e) (f)
O
Figure 5.27: Scenario 3: the multi-robot case in which the third robot failsbefore completing its task
If the failure of the third occurs before completing its assigned task, then arecovery is needed (Figure 5.27). Using Plan B Precaution Routines, DEMiR-CF
in small instances. However, these approaches may become impractical when
the number of targets increases or the distances change frequently because
of uncertain knowledge, the dynamism of the environment or the changing
structure of the mission. Therefore, the IP approach may be too expensive when
added to the path planning calculations for large target sets. As the resultsillustrate, allocating all targets and generating routes of robots from scratch may
result in highly suboptimal solutions.
DEMiR-CF eliminates the redundant efforts by means of the incremental
assignments based on the up-to-date situations of the environment for this
domain. It can also handle the contingencies by Plan B Precaution Routines.
Communication failures may sometimes prevent target allocations from being
optimal. The framework can also detect these situations and maintains highsolution quality by the dynamic task selection and task exchange scheme. As a
final remark, as our experiments and given sample situations reveal, we argue
that target allocation and route construction should be integrated for better
results in this domain. This integration and incremental allocation is useful for
eliminating redundant calculations in highly dynamic or unknown environments.
In this chapter, we evaluate the performance of DEMiR-CF in Naval Mine
CounterMeasure (MCM) missions and Naval Homeland Security Missions.
Marine applications using AUVs (Autonomous Underwater Vehicle) involve
challenges in addition to noisy communication, position uncertainty and the
likelihood of failures. In particular, in undersea operations communication
windows are restricted and bandwidth is limited. Consequently, coordination
among agents is correspondingly more difficult. It is highly likely that the
initial task assignments are subject to change during run time in these kinds
of environments. DEMiR-CF on NAVY missions can ensure robust execution
and efficient completion of missions against several different types of failures.
The experiments are performed on the US Navy’s Autonomous Littoral Warfare
Systems Evaluator-Monte Carlo (ALWSE-MC) simulator against different
contingencies that may arise at run time.
Empirical evaluations are performed on a cooperative NAVY mine clearancemission (Sariel et al., 2006b) and a cooperative NAVY homeland security mission
(Sariel and Balch, 2006b).
6.1. Naval Mine Countermeasure Missions
Naval Mine CounterMeasures are actions taken to counter the effectiveness of
underwater mines. MCM operations include finding and seizing mine stockpiles
before they are deployed, sweeping desired operational areas, identifying mined
areas to be avoided, and locating and neutralizing individual mines (Stack and
Manning, 2004).
Our research is focused on the subset of MCMs that involve locating and mapping
all individual mines in an operational area. In general, recognizing proud mines
on the seafloor is not overly difficult; the difficulty arises with the abundance
of non-mine objects on the sea floor that possess mine-like characteristics (e.g.,
geologic outcroppings, coral, man-made debris, etc.). This ample supply of
false alarms has necessitated the following strategy typically employed by theNavy: detect and classify the mine-like objects (MLOs) with high-coverage rate
sensors (e.g., sidelooking sonar), employ advanced signal processing techniques
for maximal false alarm reduction, then revisit the remaining MLOs with
identification-quality assets (e.g., electro-optic sensors) to confirm them as mines
or dismiss them as false alarms.
The reference mission in this research is to detect, classify, and identify
underwater mines in a given operational area simulated in a PC-based
software, ALWSE-MC (ALWSE), analysis package designed to simulate multiple
autonomous vehicles performing missions in the littoral regions including mine
reconnaissance, mapping, surveillance, and clearance. This mission employs
two types of vehicles: unmanned underwater vehicles (UUVs) which are free
swimming AUVs and possess large-footprint sensors (e.g., side-scan sonar) for
detection and classification (D/C) of mines and sea-floor crawlers equipped withshort-range, identification-quality sensors (e.g., camera). The crawlers have the
ability to stop at an object and take a picture with a camera.
6.2. Applying DEMiR-CF to the Naval MCM
To apply DEMiR-CF framework to the NAVY MCM, the task types and the
vehicle operations are determined and the general representation is adopted to
be used in this domain. Different operations are needed for different underwater
vehicles depending on their capabilities and the task types that they can execute.
6.2.1. Task Representation for The MCM
Our general task representation is capable of describing complex tasks with
interdependencies. However, in this particular case study, tasks do not have
interdependencies. Two types of tasks are defined for vehicles: “visit waypoint”
(w) and “identify MLO” (t). The task representation includes the capabilities
required for each type of task: reqcapw contains side-scan sonar and reqcapt
contains cameras besides the standard capabilities of AUVs common in both typesof vehicles. The coverage mission (M C ) contains a predefined number of known
waypoints (wi ∈ M C , 0 < i ≤ ||M C ||) to be visited by all UUVs (RU U V ⊂ R).
One way of task representation is to directly assign a task for each waypoint.
However, this representation has a drawback of high communication requirements
for the efficient completion of the mission. Instead, we represent tasks as interest
points of regions/search areas (W k = ∪wi, ∀wi is unvisited, and W k ⊆ M C ).
Therefore, both the allocation of the waypoints to the robots and the paths
constructed to traverse these waypoints are determined online by negotiations.Negotiating the interest points (regions) instead of the individual waypoints
reduces the communication overhead. Regions determined by different UUVs
may vary during runtime and may sometimes overlap. However, the uncertainty
related to the region determination is within an acceptable range, especially
when the cost is compared to the requirements of complete knowledge sharing
by representing each waypoint as a task. Before defining the regions, the relativedistance values, reldist(r j, wi), are determined for each unvisited waypoint wi
using Equation 6.1, where function dist returns the Euclidean distance between
points. rk locations are the latest updated locations of the robots. If there is no
known active robot assumed to be running properly, reldist(r j, wi) is the value
of the distance between the robot and the waypoint.
reldist(r j , wi) = dist(r j, wi) − min∀k= j (dist(rk, wi)), rk is active (6.1)
Each robot defines its rough regions (W jk , 1 ≤ k ≤ ||RU U V ||). The number of
regions equals the number of UUVs believed to be running properly. After
sorting the reldist(r j , wi) values of the unvisited waypoints in descending order
as an array, the array is cut into subarrays which represent the regions. Each
region contains approximately an equal number of waypoints. Each robot
specifies the region of the highest interest as its “first” region. If the robots
are closely located, the regions of highest interest may overlap. In this case,
negotiations are needed to resolve conflicts and to assign only one robot for each
region.
The identification mission (M I ) contains an unknown number of tasks for the
MLO locations (ti ∈ M I , 0 < i ≤ ||M I ||) to be visited by the crawlers. Therefore,
the tasks in M I are generated during runtime.
6.2.2. Exploration for Detection and Classification of MLO Locations
To begin the mission, the UUVs survey the operational area following waypoints
determined a priori ; however, corresponding regions containing waypoints may bereassigned by the negotiations among UUVs autonomously. After determining the
regions, each UUV proposes an auction for the region of highest interest (interest
point). After negotiations on several auctions, each UUV is assigned to the closest
region (interest point). If more than one robot is almost the same distance from
the interest point, the one with the smaller id number is assigned to the region.
The other UUVs continue to offer auctions for the remaining regions. Allocations
of the regions may also change during run time to maintain the solution quality.
Whenever UUVs detect UUV failures or recoveries from failures, they changetheir region definitions accordingly and offer new auctions. After the region
This heuristic function considers boundary targets, wb1 and wb2 in W j , the
targets with the maximum distance value. The basic idea of this function is to
forward the robot to one of these boundary targets since these targets determine
the diameter of the region (W j ) and both of them should be visited. If the robot
initially heads towards one of the boundary targets, the diameter (the longest
path) can be traveled by visiting other targets along the path. The cost penaltyapplied to forward the robot to the boundary targets is limited to a small degree.
By introducing a constant (α), this degree of direction can be adjusted. When
α is assigned a value of 2/3, this heuristic function produces close to optimal
results for multi-robot multi-target domains as explained in the previous chapter.
If more than one pair of boundary targets exist, the pair which has a member
the smallest distance from the UUV is selected.
As UUVs detect the MLOs on their way, they broadcast these estimated targetpositions to all AUVs (i.e., tasks for crawlers are generated online during
execution). Then MLO information can propagate in bucket-brigade style to
all other AUVs in the group that can possibly be reached. Periodic broadcasting
of important information coming from either owned sensors or external agents is
a way to handle communication range limitations.
6.2.3. Identification of Mine-like Objects
When the crawlers are informed about the MLO locations, they update theirworld knowledge and dynamically select the best MLO targets to visit and
propose auctions. Therefore, they can switch among tasks when new tasks
appear, if it is more profitable. It is also possible that a crawler may inadvertently
discover a mine without being informed of its position by a UUV. In this case,
the crawler identifies the target, adds it to its task list as an achieved task,
and broadcasts achievement information to maintain the system consistency.
Crawlers determine the bid values by using the MTRP cost evaluations to model
In the identification task, when crawlers are within an area close to an MLO
location, they begin keeping time while surveying the MLO location. Whenever
the time limit is reached, they set the task status as achieved and broadcast
this information. If a detection event occurs during this time period, the MLO
location is considered to be an actual mine and the task is assumed to beachieved; otherwise, it is determined as a false alarm after deadline. In either
case, the task is marked as achieved.
A conceptual flowchart summarizing operations of UUVs and crawlers, and the
general operations implemented by both types of AUVs is given in Figure 6.1.
failures, recoveries, owninconsistencies, new MLO tasks
System Model is updated/ Mission
Execution Begins
failures, recoveries , owninconsistencies, new tasks
no change
Figure 6.1: Conceptual flowchart related to the AUV operations
6.3. Experimental Results on the Naval MCM
The performance of our framework and the precaution routines is evaluated in
ALWSE-MC. Three sample scenarios in the simulator are given to illustrate the
performance for Naval MCM missions. UUVs are equipped with sensors capable
of detecting mines within 30 feet from the skin of a target in the simulator.However, they are not able to correctly identify them. The crawlers are equipped
with cameras which can both detect and identify mines within 20 feet. None of
the AUVs have predefined search patterns. UUVs have internal navigation errors,
therefore, their estimated location values are different from actual locations in
most cases. Two AUVs can communicate with each other whenever the receiver
AUV is in the sender AUV’s transmitter range, within its transmitter beamwidth, and the sender AUV is within the transmitter AUV’s receiver beam width.
All UUVs and crawlers begin execution from a deployment area. There is no a
priori information about mine locations. 121 waypoint locations (environment
size: 200x200) are known but are not assigned initially. UUVs begin negotiations
and divide the overall mission area into three (known number of UUVs)
regions. Since they are within line of sight, they can communicate their location
information. Therefore, initially defined regions are almost the same for allUUVs. Figure 6.2 illustrates a successful mission scenario with three UUVs
and two crawlers. The legend for all simulation scenarios are also provided in
the figure. Allocations of waypoints after negotiations can be seen in Figure
6.2(b). Since there are no failures, the waypoint assignments do not change
during run time. However, the crawlers sometimes switch among tasks if they
are not informed about tasks that are being executed and sometimes parallel
executions occur. Whenever they are in communication range, they can resolve
the conflicts efficiently by means of the precaution routines. As in Figure 6.2(a),
the crawlers can also detect mines without being informed (red circled in the
figure). The routes of the crawlers may seem somewhat random. However, it
should be noted that the tasks related to the MLO locations appear during run
time when they are discovered, and the communication range is limited.
UUV1UUV2
UUV3UUV4
Crawler1Crawler2
Mines
Detection by crawlersUUV 1 Search Area
UUV 2 Search Area
UUV 3 Search AreaUUV 4 Search Area
(a) (b)
Figure 6.2: Scenario 1: (a) The UUVs cover the area; the crawlers visit MLOlocations.
In the second scenario, one of the UUVs (UUV3) fails in the same setting of scenario 1 (Figure 6.3). Initially all UUVs begin execution (Figure 6.3(a)).
range, UUV4 redefines its regions by considering incoming information for the
visited waypoints. Videos of these scenarios are available at the website-ref:
SarielMCMMov.
(a) (b)
(c) (d)
(e)
Figure 6.4: Scenario 3: UUV failure is handled and new robot arrival is alsoused to improve the system utility in a distributed manner
In the same settings, another experiment is conducted to evaluate the effect of
message loss rate on the mission completion success. Table 6.1 illustrates the
results (µ | σ) averaged over 10 runs. When the message loss rate is differentfrom 0, as expected, the mission completion time performance of the system
the generated task instances (Figure 6.6(a)). The robots patrol the subareas
which are determined after the negotiations (Figure 6.6(b)). Therefore, although
there is only one task on the higher level, the robots create instances of the
Search Task (Search 1-3) as if each instance is another separate task. If there
are no hostile intentions, the robots only search the area.
Search-1 Search - 3
SearchS T
[reqno = 1]
(a) Mission Graph
Search - 2
SearchS T
[reqno = 1]
(b) Allocation of the Mission Tasks
R1
R2
R3
Figure 6.5: Initial mission graph consists of only the search task
−50 0 50 100 150 200 250−50
0
50
100
150
200
250
Robot Paths and Search Areas
SA−1 SA−2 SA−3 R1 R2 R3
−50 0 50 100 150 200 250−50
0
50
100
150
200
250Robot Paths and Search Areas
(a) (b)
Figure 6.6: Robots patrol the area in the corresponding regions
Whenever a hostile diver is detected by the robots, a related interception task is
generated. The scenario is illustrated in Figure 6.7. The robots begin searching
the area (Figure 6.7(a)). R2 recognizes the hostile intent (Figure 6.7(b)). After
detection, the hostile vehicle attacks R2. R2 returns to the deployment ship.
R1 takes control of the intercept task. The hostile intention disappears (Figure
6.7(c)). R1 and R3 continue searching the area (Figure 6.7(d)).
The updated mission graph for the sample scenario is illustrated in Figure 6.8.The hostile diver may be destructive by using missiles. Therefore, task execution
Figure 6.7: A sample execution trace under highly dynamic task situations inwhich failures occur after shots by the hostile diver
may need to be preempted and the task execution authority is exchanged during
run time. The robots may need to generate local tasks (e.g., Repair/Refuel Task,
which is generated by R2 after being shot by the hostile diver unexpectedly) as
in Figure 6.8(d).Therefore, the mission graphs may be different for the robots
even when they work cooperatively (Figure 6.8(c-d)). In Figure 6.8(c), although
executing the Intercept Task, R1 can make a coalition commitment assuming it
will succeed in a predefined time period (described as TBD). At this time R2
cannot make any coalition commitment for the search task because its future
operations depend on its recovery.
Cost evaluation for the tasks are computed by considering the task facilitating
composite (multi) objective missions. While the robots try to optimize the fuel
levels for the search task, the intercept task requires immediate response and time
minimization. Therefore, different cost evaluations are carried out for different
tasks. We provide the cost evaluations for the task types used in the experiments
in Table 6.2. Cost evaluation for the search task is implemented by first dividing
the search area into regions (with corresponding interest points) and comparing
the distance values for these interest points. The same cost evaluation used in
the MCM mission is adopted for the search task. For the intercept task, theexpected time to achieve (intercept the diver) the task is taken as the cost value.
In this chapter, we investigate the performance of DEMiR-CF on complex
missions with resource constrained and interrelated tasks. Different from
previous chapters, the robots take part in more complex tasks where they
interact with the objects in the environment (Sariel et al., 2007a). The objective
is not only optimizing cost functions but also obeying rules and resolving
constraints on task execution during runtime. When the tasks of a mission are
interrelated and subject to several resource constraints, more efforts are needed
to coordinate robots towards achieving the mission than during independent
tasks.
As in the previous experimental evaluations, the DEMiR-CF framework is
evaluated for complex domains. The incremental task selection and allocation
mechanisms of DEMiR-CF also eliminate redundant considerations in this
domain. The base mechanisms of DEMiR-CF are used for designing the solution.
Rough schedule formation and cost evaluations are designed according to thecomplex mission requirements.
7.1. Complex Multi-Robot Mission Problem Statement
The multi-robot task allocation problem is better viewed as a scheduling problem
if there are interrelations among tasks, suggesting the use of Operation Research
methods. However, when the problem solving time is limited and/or reallocations
are frequently required at runtime, Operation Research (OR) solutions such
as Branch and Bound (Brucker et al., 1998) or Integer Programming methods
may not be directly applicable. In this chapter, our focus is on complex
missions (project tasks) with interrelated tasks whose requirements on task
execution may vary. These interrelations may correspond to shared resources,
producer/consumer, simultaneity and task-subtask dependencies (Ossowski,
1999). The Pick-Up/Delivery domain tasks can be classified in this class
because of the producer/consumer type of dependency relation for the pick-up
and delivery tasks. More complicated interrelations may be placed in mission
representations. Simultaneous execution requirements imply tightly coupled taskexecution where the actions implemented by each robot are highly dependent on
the actions of others. Furthermore, a group of robots executing a task may be
either homogeneous or heterogeneous. The heterogeneity may be in the possessed
capabilities or in the task execution performance. For example, robots may have
the same equipments capable of achieving all the tasks of the mission but may
differ in abilities such as speed. According to the classification of multi-agentorganizations given in Horling and Lesser (2005), coalitions (agent groups) are
formed to perform tasks in cooperation. From our perspective, coalitions are
suitable for meeting the simultaneous resource requirements of executing tasks
with a subteam of robots.
7.2. Applying DEMiR-CF to Complex Missions
The dynamic and incremental task selection, distributed allocation and
contingency handling mechanisms of DEMiR-CF are used in the design of the
solution for complex mission achievement. This domain forms a platform to
apply the full functionality of DEMiR-CF.
Although the base mechanisms are the same, the rough schedule generation
scheme and the cost functions are designed accordingly to meet the interrelation
and resource constraints.
7.2.1. Dynamic Priority-based Task Selection Scheme and OnlineScheduling
As the core principle of DEMiR-CF, the robots make instantaneous decisions
(from their local perspectives) which are both precedence and resource feasible
in the context of the global-time extended view of the problem. While the
completion of the mission is the highest priority objective, performance related
objectives can additionally be targeted. Each robot initially forms a rough
schedule of its activities for an overall time extended resolution of the mission.
Since these schedules are highly probable to change in dynamic environmentsand the robots also have the real-time burdens of path planning, mapping etc.,
the rough schedules formed are tentative and constructed by computationally
cheap methods (explained in subsection 7.2.3.). Therefore, the robots in our
framework come up with their rough schedules and refine their plans during
actual fast execution when information available in the current context enables
them to make specific, detailed decisions.
Instead of scheduling all tasks in one step, we propose a Dynamic Priority-basedTask Selection Scheme (DPTSS) to allocate tasks to robots incrementally,
considering the global solution quality. The main objective of the proposed
scheme is the incremental allocation of tasks by taking into account the
precedence and resource constraints whenever a new task needs to be assigned,
instead of scheduling all tasks from scratch.
The CTSP , introduced earlier, is an optimization problem as in ScP and it is
desirable to find a solution by considering the problem from a global perspective.
Therefore, the instantaneous task selection scheme needs to be strengthened by
considering the problem as a whole, with the designed cost evaluation functions.
Depending on the objective function, either priorities or penalties can be applied
to find an efficient solution, ensuring a time-extended view of the problem.
Each robot r j generates its rough schedule as a dynamic priority queue similarto runqueues by considering its critical task list (LCj ), the eligible task set
(T Ej ), the conjunctive arcs (if any) and the requirements. If there are no new
online tasks or invalidations, the order of the tasks which are connected by the
conjunctive arcs remains the same in the priority queue, even though there may
be additional intermediate entries into the queue at runtime.
The critical tasks may be determined by either negotiations or beliefs. To
eliminate intractable communication overhead, we use a rough belief update
approach to form the critical tasks. Each critical task is assigned a probability
value to indicate its criticality. Critical task information is used for determining
the task requirements such as power, fuel etc.
Algorithm 7 GeneratePriorityList for robot r j
input: Eligible task set (T Ej ), active task set (T Aj)output: Topologically ordered and prioritized schedule list: S Rj
S Rj = φ, S Temp = φS Temp = DF S (T E j) /*List generated by a depth-first search, the tasks are orderedby ascending order of estimated task completion times*/for all ti ∈ S Temp do
if ti ∈ T Aj then
insert ti in S Rj as ordered by the cost value and the precedenceelse
insert ti to the front of S Rj
end if
end for
The rough schedule of a robot constitutes a topological order of the directed
rough schedule of the robot may be empty. In this case, the robot selects to stay
idle as determined in the DPTSS algorithm.
7.2.1.1. DPTSS Algorithm
In our incremental allocation approach, the fundamental decision that each robotmust make is the selection of the most suitable task from the active task set ( T A)
by considering the eligible task set (T E ). Algorithm 9 presents the DPTSS in
which a rough schedule is generated before making a decision. The four different
decisions made by robots after performing the DPTSS are:
• continue to execute the current task (if any),
• join a coalition,
• form a new coalition to perform an available task or
• stay idle.
Algorithm 9 DPTSS Algorithm for robot r j
input: Mission (M ) task descriptionsoutput: Action to be performed depending on the selected task
Determine the T Ej , T Aj ⊆ T Ej and LCj ⊆ T Ej
/*GenerateListOfCriticalTasks*/LCj = φ
for each ti ∈ T Ej do
P ct(ti) = reqno#ofsuitablerobots
if P ct(ti) ≥ 0.5 then
Insert ti in LCj prioritized by the P ct(ti)end if
end for
[S Rj, ts] = GenerateRoughSchedule (T Ej , T Aj , LCj , curcs j)
if ts = φ thenif ts is the current task then
Continue with the current executionelse
Offer an auction to form a new coalition or directly begin executionend if
else
if ts ∈ T ie and it is profitable to join the coalition then
The dynamic task switching scheme is used by robots to dynamically switch
between tasks if updates in the world knowledge compel. Therefore, issues
related to both online scheduling and scheduling under uncertainty are addressed.
The DPTSS process is repeated whenever a robot completes its current taskexecution or detects a change in its world knowledge. Instead of regenerating the
rough schedule at each call of the DPTSS, the rough schedule may be repaired
whenever it is desirable.
7.2.2. Distributed Task Allocation Scheme
Standard auction procedures of our distributed allocation procedures are applied.
For task executions with multiple robot requirements (for which reqno > 1),
coalitions are formed. For such a task, a coalition leader and the required numberof coalition members are selected. These roles are assigned to ensure synchronous
task execution among coalition members although tight coordination routines
are beyond the scope of this research. Additionally, precaution routines are
added to check validity, consistency and coherence in these negotiation steps.
Each robot intending to execute a task announces an auction after determining
its rough schedule and performing the DPTSS. Basically, auction announcements
are ways to illustrate intentions to execute tasks for which reqno = 1 or to select
members of coalitions to execute tasks for which reqno > 1. The succeeding
steps of the distributed task allocation scheme are applied as in the general
design of the framework.
7.2.3. Cost/Bid Evaluation and Tie Breaking Rules
The cost evaluation has a tremendous impact on the solution quality. Each task
type as a part of the mission requires a different cost evaluation to efficiently
solve the problem. As an example, while estimating the cost value for picking
up an item, the distance between the robot location and the estimated location
of the item may be considered. However, to form a globally efficient allocation,
the locations of the other items must be considered as well. We have validated
this statement in the MTRP domain. Cost evaluation is performed by using the
corresponding functions given in Table 7.1. If a robot is executing a task when
it receives an auction message, it sends the bid value by considering the final
destination of the current task as the location of itself.
Another important design criteria is determining the bid values to be sent to the
Locate/Pick-up Estimated time to reach at the location of the object.Deliver/Push Estimated time to carry/push the object from the
initial location to the final destination.Clean Estimated time to cover the whole environment.
other robots. Furthermore, robots may need to cancel or postpone their offers
for auctions if there is synergy between tasks when announcements of offers from
others are received. Intuitively, robots only send bid values for the active tasks
for themselves.
A common situation appears when the auctions are offered at the same timeby different robots either for the same task or for different tasks. To decide on
the selection of the robots to execute specific tasks in a distributed setting is a
challenging issue. In our approach, if there are conflicting auctions for the same
task, only the robot offering an auction with the smallest cost value continues with
the auction negotiation process. In the case of the conflicting auctions for different
tasks, a resource-based rule (related to the reqno of the tasks) borrowed from
Operations Research, Greatest Resource Requirements (GRR), is used (Brucker
and Knust, 2006).
7.2.4. Analysis of the Approach
The approach we have proposed for this particular problem domain offers a
polynomial time solution. The critical task list generation takes O(nlog(n)) time
for all n number of tasks. Achievability of the mission is determined in O(n).
The complexity of the rough schedule generation is bounded by the topological
list generation algorithm which is in the order of O(n + e) (where e is the number
of conjunctive arcs, i.e., hard dependencies). Therefore, the total complexity
becomes O(n(e+nlog(n))). If (e << n), the complexity of the proposed approach
reduces to O(n2log(n)).
7.3. Complex Mission Experiments
We have conducted real-world experiments and real-time dynamic simulation
experiments on WEBOTS, the professional mobile robot simulation software. It
contains a rapid prototyping tool to create 3D virtual worlds with robots andobjects possessing physics properties (WEBOTS).
In our simulation experiments, each environment is represented as a 5m by 5m
3D virtual world where 70mm-size simulated Khepera II robots and objects are
located. The environments are randomly generated VRML files containing the
robots and the objects. Each Khepera II robot is mainly equipped with a 25MHzMC68331 micro-controller, 512K Flash and 512K RAM memories and 8 infra-red
sensors with a limited obstacle detection range as simulated in WEBOTS.
Communication is achieved through wireless links in both simulations and in
real-world experiments. Real Kheperas have standard radio turrets mounted
on them to communicate through the selected radio frequency. Emitters in the
simulator are configured to have a baud rate of 9600 and a buffer size of 1024B
as in the receiver modules.
A multi-process controller implementation is performed on the real robots to
achieve the proposed system. Different modules on the task allocation layer are
integrated with the lower-level Sensory Interface, Motor Interface, Motion Model
and Mapping modules in the multi-threaded structure.
0 2 4 6 8 10 12
0
500
1000
1500
2000
2500
3000
3500
4000
Number of Robots
T i m e ( s )
Figure 7.1: Mission completion time(s) for the pick-up/delivery mission, withtask number fixed at 20
The first set of experiments is targeted to analyze the scalability of the proposed
approach on the pick-up/delivery mission in which the tasks are interrelatedby picking up and delivery constraints. All picked up items are collected
Figure 7.2: Total path length traversed by the robots(mm) for thepick-up/delivery mission, with task number fixed at 20
in the center of the environment. In the fully flexible version, while there
are precedence constraints between pick-up and delivery tasks, there are no
interrelations between pick-up tasks for different items. The items are distributed
in the environment at fixed locations for each run. The robot locations are
randomly determined. Figure 7.1 illustrates the mission completion times for
sets with different numbers of robots. As expected from this approach, the time
to complete the overall mission is greatly reduced with increasing numbers of
robots, validating the scalability of the approach. Figure 7.2 plots the total path
length traversed by the robots. Since the items are delivered to the center of
the environment, an extreme variation for the expected utility in the total path
length traversed by robots is not expected, as illustrated in the graph.
A sample scenario for a complex mission which includes tasks for pushing boxesand picking-up and delivering items to a desired location is given in Figure 7.3
with five participating robots. Locations of the objects are indicated with the
red arrows in the figure. In the first scenario, two items are picked up and
delivered to the destinations by the robots possessing grippers. The two robots
simultaneously and independently push the two boxes. One of the robots stays
idle during mission execution. In the second scenario, since the minimum required
number of robots to push one of the boxes is two, the two robots form a coalition
Figure 7.3: Scenario 1 and 2: Robots push and carry boxes to a final destination.Some tasks may require simultaneous and tightly coordinated task execution.
Another complex mission allocation scenario which includes tasks for pushing
a box, carrying a cylindrical object to a final destination and then inspecting
the environment is implemented by three Khepera II robots and the execution
scenario is illustrated in Figure 7.4. There are interrelations between push,
carry and inspection tasks respectively as in the graph depicted in Figure
7.5. While the objects can only be carried by the robots with grippers, the
inspection task requires possessing a camera. The box can be pushed by any
of the three robots. However, due to the cost evaluations and the critical task
list consideration, allocations are implemented accordingly. Robots obey the
interrelation constraints and each robot takes part in a suitable task execution
for itself in which the decision is made in a completely distributed manner. The
video of this scenario is available at the website-ref: SarielKh2Mov.
In the second set of experiments, the task allocation approach of DEMiR-CF is
compared with both the Prim Allocation approach and the Integer Programming
approach. As expected, both DEMiR-CF and the Prim Allocation methods have
tractable computational complexities compared to the Integer Programming
approach. However, as the results presented in Section 5. reveal, DEMiR-CFintegrated with new heuristic cost functions produces results that are close to
the optima for the multi-robot case of the problem.
In the third set of experiments, the performance of DEMiR-CF under real-time,
dynamic conditions is evaluated. Online task handling scenarios in environments
where map information is not available are presented. It has been shown that,
even when the map information is not available, robots can efficiently reconfigure
themselves according to changes in the environment and incrementally selecttasks suitable for themselves through decisions that take into account the most
recent information.
In the fourth set of the experiments, real robot scenarios are presented against
robot failures to validate the robustness of DEMiR-CF.
The results of the experiments support our thesis that in a dynamic execution
environment an incremental task selection approach eliminates redundant efforts
that are introduced by allocating all tasks from scratch if there is an unexpected
change. The rough schedule generation scheme forms loose commitments, which
if needed, can be canceled in the future. Thus, it offers a way to reconsider
the problem globally when it is appropriate. The approach is efficient with its
polynomial computational and communication complexities. Plan B precautions
ensure that the CMAP is successfully solved at the end of mission execution. If
real resources permit, failures are handled to maintain system consistency. Even
though a certain amount of additional communication overhead is injected into
the system by the Plan B Precaution Routines, communication efficiency is alsoachieved as experimental results illustrate.
DEMiR-CF is also evaluated in NAVY domains where the cooperation of
underwater vehicles is required for homeland security missions, such as the mine
countermeasure mission. This domain has a challenging structure since it is
performed underwater and communication is achieved through acoustic modems.
In this domain, the robot team is modeled as a heterogeneous team and the
mission consists of different types of tasks, requiring different vehicles. Thedomain is modeled to include both the coverage problem and the MTRP.
The coverage problem is solved by having robots generate rough schedules for
the regions to be covered and negotiate over these regions. While covering the
environment, robots simultaneously sense “mine like objects” to generate new
online tasks. They transmit this information to different types of robots thatcan visit targets and perform correct classification. With this mode of operation,
the problem turns into the MTRP and is solved in a decentralized way for
continually generated online tasks. Depending on the types of robots, different
types of rough schedules are generated by either coverage or classification robots.
Regions are represented as rough schedules for the coverage tasks, whereas target
sets are formed as rough schedules for the classification tasks.
The online task handling performance of DEMiR-CF is validated throughexperiments. The robustness of the framework against both communication and
robot failures and the efficiency with which it responds to dynamic changes in
the environment are tested in simulations.
As a final domain, object construction domain is selected to use and validate
the full functionality of DEMiR-CF. Complex missions investigated in this
domain involve tasks with resource constraints and interrelations. Multi-robot
task allocation problem is better viewed as a scheduling problem if there are
interrelations among tasks. Therefore, this domain forms the basis of the design
objective of DEMiR-CF. A topological list generation approach for forming
rough schedules is used to solve the CTSP and to resolve the constraints on
tasks. Applicable cost functions are proposed for separate tasks for robots to
achieve an overall complex mission. The scalability of DEMiR-CF is validated
through experiments. Efficiency is analyzed and validated both in theory
and in experiments. As results illustrate, the CTSP is solved by each robot
implementing the incremental task selection, distributed task allocation and
contingency handling mechanisms of DEMiR-CF to achieve the overall CMAP
for complex missions.
In conclusion, in this PhD thesis, DEMiR-CF has been designed and implemented
as a generalized framework for cooperative multi-robot mission achievement.
Several performance tests are carried out for different domain implementations
of the framework. It is demonstrated that the framework is an efficient,
complete, scalable and robust decentralized framework for a multi-robot system.
Furthermore, DEMiR-CF is shown to be applicable on even very small andsimple robots with cheap computational capabilities, such as Khepera II.
If embedding a commercial program such as CPLEX IP solver on robots
is possible and the requirements of the decision frequency and change in
task allocations do not affect the response time of the system, an Integer
Programming method may be used to generate rough schedules of robots.
In this case, redundant allocations can be implemented to reach globally
optimum results for the current situation. This approach can also be successfully
integrated to the incremental assignment approach that we propose. However,
as we mentioned earlier, if there are computational limitations on the robot
hardware, a heuristic cost evaluation is inevitable for the system.
In the current design of the DEMiR-CF, planning activity is performed for the
MTRP domain. Route construction is a high-level path planning problem and is
solved by DEMiR-CF by integrating it with task allocation. For more complex
tasks, a global plan with interrelations between tasks is given to the robots
initially. DEMiR-CF is capable of changing the structure of given plans during
runtime in a decentralized way. Online tasks are integrated into the task graph.
However, this research area desires more investigation for both constructing a
global plan by the robots and satisfying and resolving conflicts in the global planas investigated in MPCP.
There are no previously designed test-beds for multi-robot systems research.
Although comparisons and results are provided, usually it is hard to evaluate
and compare implemented systems with the insufficient implementation details
presented in the publications. Even though RoboCup leagues present test-beds
for the comparison of the architectures, research papers can only present ad-hoc
implementations without comparisons for real robots. Therefore, a formalismand designs of test-beds are greatly needed to improve the field and find better
Gancet, J., Hattanberger, G., Alami, R. and Lacroix, S., 2005. TaskPlanning and Control for a Multi-UAV System: Architecture andAlgorithms. In IEEE/RSJ Intl. Conf. on Intelligent Robots andSystems (IROS).
Gerkey, B. and Mataric, M. J., 2002. Sold!: Auction Methods for MultirobotCoordination. IEEE Trans. Robot. Automat., 18(5), 758–768.
Gerkey, B. and Mataric, M. J., 2004. A Formal Analysis and Taxonomy of Task Allocation. Intl. J. of Robotics Research, 23(9), 939–954.
Ghiani, G., Guerriero, F., Laporte, G. and Musmanno, R., 2003.Real-Time Vehicle Routing: Solution Concepts, Algorithms andParallel Computing Strategies. European Journal of Operational Research, 151, 1–11.
Goldberg, D., Cicirello, V., Dias, M. B., Simmons, R., Smith, S.,Smith, T. and Stentz, A., 2002. A Distributed LayeredArchitecture for Mobile Robot Coordination: Application to SpaceExploration. In 3rd Intl. NASA Workshop on Planning andScheduling for Space .
Goldberg, D., Cicirello, V., Dias, M. B., Simmons, R., Smith, S.,Smith, T. and Stentz, A., 2003. Market-Based Multi-RobotPlanning in a Distributed Layered Architecture. In Multi-RobotSystems: Swarms to Intelligent Automata, Proc. Of Intl. Workshopon Multi-Robot Systems .
Hazon, N. and Kaminka, G. A., 2005. Redundancy, Efficiency, andRobustness in Multi-Robot Coverage. In IEEE Intl. Conf. onRobotics and Automation (ICRA).
Horling, B. and Lesser, V., 2005. A Survey of Multi-Agent OrganizationalParadigms. The Knowledge Engineering Review , 19(4), 281–316.
Huhns, M. N. and Stephens, L. M., 1999. Multiagent Systems: A ModernApproach to Distributed Artificial Intelligence, chapter MultiagentSystems and Societies of Agents, pages 80–120. The MIT Press.
Jarnik, V., 1930. O jistem problemu minimalnim (About a certain minimalproblem). Prace Moravske Prirodovedecke Spolecnosti, 6, 57–63.
Jennings, J. J., Whelan, G. and Evans, W. F., 1997. CooperativeSearch and Rescue with a Team of Mobile Robots. In International
Conference on Advanced Robotics (ICAR).Jennings, N. R., 1996. Foundations of Distributed Artificial Intelligence,
chapter Coordination Techniques for Distributed ArtificialIntelligence, pages 187–210. John Wiley and Sons.
Kalra, N., Ferguson, D. and Stentz, A., 2005. Hoptiles: A Market-BasedFramework for Planned Tight Coordination in MultiRobot Teams.In IEEE Intl. Conf. on Robotics and Automation (ICRA).
Kaminka, G. A. and Tambe, M., 2000. Robust Multi-Agent Teams viaSocially-Attentive Monitoring. Journal of Artificial Intelligence Research, 12, 105–147.
Kitano, H., 2000. Robocup rescue: A grand challenge for multi-agent systems.In Proceedings of ICMAS .
Koes, M., Nourbakhsh, I. and Sycara, K., 2005. Heterogeneous MultirobotCoordination with Spatial and Temporal Constraints. In AAAI National Conference on Artificial Intelligence .
Kose, H., Kaplan, K., Mericli, C., Tatlidede, U. and Akin, L.,2005. Market-Driven Multi-Agent Collaboration in Robot Soccer
Lagoudakis, M. G., Berhault, M., Koenig, S., Keskinocak, P.and Kleywegt, A., 2004. Simple Auctions with PerformanceGuarantees for Multi-Robot Task Allocation. In IEEE/RSJ Intl.Conf. on Intelligent Robots and Systems (IROS).
Lagoudakis, M. G., Markakis, E., Kempe, D., Keskinocak, P.,Kleywegt, A., Koenig, S., Tovey, C., Meyerson, A. andJain, S., 2005. Auction-Based Multi-Robot Routing. In Robotics:
Science and Systems (RSS).
Lawler, E. L., Lenstra, J. K., Kan, R. and Shmoys, D. B., 1985. TheTraveling Salesman Problem: A Guided Tour of CombinatorialOptimization. John Wiley & Sons, New York, NY.
Lemaire, T., Alami, R. and Lacroix, S., 2004. A Distributed Task AllocationScheme in Multi-UAV Context. In IEEE Intl. Conf. on Robotics and Automation (ICRA).
Malone, T. W. and Crowston, K., 1994. The Interdisciplinary Study of
Sandholm, T. W., 1999. Multiagent Systems: A Modern Approach toDistributed Artificial Intelligence, chapter Distributed RationalDecision Making, pages 202–258. The MIT Press.
Sariel, S. and Akin, H. L., 2005. A Novel Search Strategy for Autonomous
Search and Rescue Robots. In RoboCup, volume 3276 of Lecture Notes in Computer Science , pages 459–466. ISBN 3-540-25046-8.
Sariel, S. and Balch, T., 2005a. Real Time Auction Based Allocation of Tasksfor Multi-Robot Exploration Problem in Dynamic Environments.In Integrating Planning into Scheduling: Papers from the 2005 AAAI Workshop, WS-05-06 , pages 27–33.
Sariel, S. and Balch, T., 2005b. Robust Multi-Robot Coordinationin Noisy and Dangerous Environments. TechnicalReport GIT-GVU-05-17, GVU Center, Georgia Institute of
Technology.Sariel, S. and Balch, T., 2006a. Distributed Autonomous Robotic Systems
(DARS) 7, chapter A Distributed Multi-Robot CooperationFramework for Real Time Task Achievement, pages 187–196.Springer Verlag.
Sariel, S. and Balch, T., 2006b. Dynamic and Distributed Allocation of Resource Constrained Project Tasks to Robots. In Multi-AgentRobotic Systems (MARS) Workshop at the Third International Conference on Informatics in Control, Automation and Robotics .
Sariel, S. and Balch, T., 2006c. Efficient Bids on Task Allocation forMulti-Robot Exploration. In The 19th International The Florida Artificial Intelligence Research Society (FLAIRS) Conference .
Sariel, S., Balch, T. and Erdogan, N., 2006a. Robust Multi-RobotCooperation Through Dynamic Task Allocation and PrecautionRoutines. In The International Conference on Informatics inControl, Automation and Robotics (ICINCO).
Sariel, S., Balch, T. and Erdogan, N., 2007a. Incremental Multi-Robot
Task Selection for Resource Constrained and Interrelated Tasks. In IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS).
Sariel, S., Balch, T. and Stack, J. R., 2006b. Distributed AutonomousRobotic Systems (DARS) 7, chapter Empirical Evaluation of Auction-Based Coordination of AUVs in a Realistic Simulated MineCountermeasure Task, pages 197–206. Springer Verlag.
Sariel, S., Balch, T. and Stack, J. R., 2006c. Distributed Multi-AUVCoordination in Naval Mine Countermeasure Missions. TechnicalReport GIT-GVU-06-04, GVU Center, Georgia Institute of
Sariel, S., Erdogan, N. and Balch, T., 2007b. An Integrated ApproachTo Solving The Real-World Multiple Traveling Robot Problem.In The 5th International Conference on Electrical and Electronics Engineering .
Shehory, O. and Kraus, S., 1998. Methods for Task Allocation via AgentCoalition Formation. Artificial Intelligence , 101, 165–200.
Simmons, R. and Apfelbaum, D., 1998. A Task Description Language forRobot Control. In Conference on Intelligent Robotics and Systems .
Smith, R. G., 1980. The Contract Net Protocol: High Level Communicationand Control in a Distributed Problem Solver. IEEE Transaction onComputers C-, 29(12), 1104–1113.
Stack, J. and Manning, R., 2004. Increased autonomy and Cooperationin Multi-AUV Naval Mine Countermeasures. In Proceedings of Undersea Defence Technology .
Tews, A., 2001. Adaptive Multi-robot Coordination for Highly DynamicEnvironments. In International Conference On Computational Intelligence for Modelling (CIMCA).
Toth, P. and Vigo, D., 2001. The Vehicle Routing Problem. Society forIndustrial and Applied Mathematics.
Vail, D. and Veloso, M., 2003. Dynamic Multi-robot Coordination.Multi-Robot Systems: From Swarms to Intelligent Automata , 2,87–98.
Vig, L. and Adams, J. A., 2005. Issues in Multi-robot Coalition Formation.In Multi-Robot Systems. From Swarms to Intelligent Automata.Volume III , pages 15–26.
In the following heuristic function definitions, T T SP is a partial tour and k is acity not on T T SP , and {i, j} is one of the edges of T T SP .
Minimum Spanning Tree Heuristic (MST): In this method, either thePrim’s algorithm or the Kruskal’s algorithm may be used to generate the MSTfor the given set of the cities. A depth first search of T is constructed. Afterintroducing shortcuts into the depth first search, it is ensured that cities arevisited only once.
Nearest Merger Heuristic (NMH): This method corresponds to theKruskal’s minimum spanning tree algorithm. The algorithm starts with n partialtours, each of which consists of a single city, then, successively merges the toursuntil a single tour containing all cities is obtained. Each time trees to be mergedare chosen so that mincij : i ∈ T T SP and j ∈ T T SP is as small as possible.
Nearest Neighborhood Heuristic (NNH): The algorithm starts with anempty tour T. Cities k and j, for which c( j,k) minimized are found. i, j isreplaced by i, k and k, j to obtain a new tour including k. The algorithmcontinues until all cities are added to T T SP .
Nearest Insertion Heuristic (NIH): The algorithm starts with an emptytour T T SP . Cities k and j, for which c( j,k) minimized are found. i, j is the edgeof T T SP which minimizes c(i, k) + c(k, j) − c(i, j), and it is replaced by i, k andk, j to obtain a new tour including k. The algorithm continues until all cities areadded to T T SP .
Cheapest Insertion Heuristic (CIH): The algorithm starts with an emptytour T T SP . If T T SP does not include all cities, for each k, the edge i, j of T T SP
which minimizes c(T T SP , k) = c(i, k) + c(k, j) − c(i, j) is found. Then, the city k
minimizing c(T T SP
, k) is found. If i, j is the edge of T T SP
for which c(T T SP
, k) isminimized, it is replaced by i, k and k, j to obtain a new tour including k. Thealgorithm continues until all cities are added to the T T SP .
Farthest Insertion Heuristic (FIH): The algorithm starts with anempty tour T T SP . City k, which is the farthest of all the cities out of T T SP ,is inserted to T T SP . The algorithm continues until all cities are added to the T T SP .
Christofides Algorithm (CA): MST of the given set of cities is constructed.Minimum matching M ∗ for the set of all odd-degree vertices in T T SP isconstructed. An Eulerian tour for the Eulerian Graph that is the union of the
T T SP and M ∗ is found, and it is converted into a tour using the shortcuts.
Except from the FIH and the NNH, all the algorithms presented above haveworst case solutions bounded by 2 ∗ OP T . The NNH is bounded by (log n)*OPT,whereas the CA is bounded by 1.5*OPT.
TSPLIB is a library of sample instances for the TSP (and related problems) fromvarious sources and of various types (Reinelt, 1991). Each instance is given withthe coordinates of the cities to be visited. In our experiments, these instancesare used for the performance analysis, and the first node is taken as the initiallocation of the robot in each instance. While the cities are marked with red xmarkers, the initial nodes and the robot locations are indicated with blue circlesin the following graphs.
Sanem Sariel received her B.Sc. in Computer and Control Engineeringand M.Sc. in Computer Engineering from Istanbul Technical University in 1999and 2002 respectively. She has been serving as a research and teaching assistant
at ITU since 1999. She was awarded with a research scholarship to do researchfor her Ph.D. studies in the United States in 2004 where she worked with Prof.Tucker Balch in the BORG Laboratory at Georgia Institute of Technology onmulti-robot coordination. Sanem Sariel’s research interests lie in the area of Distributed Problem Solving, Multi-Robot Coordination and Intelligent Agents.The following research papers were published during her PhD research:
Book Chapters:
• S. Sariel and T. Balch,“A Distributed Multi-Robot Cooperation Frameworkfor Real Time Task Achievement”, Distributed Autonomous Robotic
Systems (DARS) 7, Springer Verlag ISBN:4431358781, 2006, pp. 187-196.
• S. Sariel, T. Balch and J. Stack, “Empirical Evaluation of Auction-BasedCoordination of AUVs in a Realistic Simulated Mine CountermeasureTask”, Distributed Autonomous Robotic Systems (DARS) 7, SpringerVerlag ISBN:4431358781, 2006, pp. 197-206.
Conference Papers:
• S. Sariel, T. Balch and N. Erdogan, “Incremental Multi-Robot TaskSelection for Resource Constrained and Interrelated Tasks”, IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS), 2007.
• S. Sariel, N. Erdogan and T. Balch, “An Integrated Approach To SolvingThe Real-World Multiple Traveling Robot Problem”, The 5th InternationalConference on Electrical and Electronics Engineering, 2007.
• S. Sariel, T. Balch and N. Erdogan, “Robust Multi Robot Cooperationthrough Dynamic Task Allocation and Precaution Routines”, The ThirdInternational Conference on Informatics in Control, Automation andRobotics (ICINCO), 2006.
• S. Sariel and T. Balch, “Efficient Bids on Task Allocation for Multi Robot
Exploration”, The Nineteenth International Florida Artificial IntelligenceResearch Society (FLAIRS) Conference, 2006.
• S. Sariel and T. Balch, “Dynamic and Distributed Allocation of ResourceConstrained Project Tasks to Robots”, Multi-Agent Robotic Systems(MARS) Workshop at the Third International Conference on Informaticsin Control, Automation and Robotics, 2006. (Also presented at the AAAIWorkshop on Auction Mechanisms for Robot Coordination, 2006).
• S. Sariel and T. Balch, “Real Time Auction Based Allocation of Tasksfor Multi-Robot Exploration Problem in Dynamic Environments”, In
Integrating Planning into Scheduling: AAAI Workshop 2005, pp.27-33.
• S. Sariel, T. Balch and J. Stack, “Distributed Multi-AUV Coordination inNaval Mine Countermeasure Missions”, GVU Tech Report GIT-GVU-06-04,2006.
• S. Sariel, T. Balch, “Robust Multi-Robot Coordination in Noisy andDangerous Environments”, GVU Tech Report GIT-GVU-05-17, 2005.
• S. Sariel, “A Framework for Multi-Robot Coordination”, The InternationalConference on Automated Planning and Scheduling (ICAPS), DoctoralConsortium, 2005.